Patients with late-stage cancer often have to endure multiple rounds of different types of treatment, which can cause unwanted side effects and may not always help.

In hopes of expanding the treatment options for those patients, MIT researchers have designed tiny particles that can be implanted at a tumor site, where they deliver two types of therapy: heat and chemotherapy.

This approach could avoid the side effects that often occur when chemotherapy is given intravenously, and the synergistic effect of the two therapies may extend the patient’s lifespan longer than giving one treatment at a time. In a study of mice, the researchers showed that this therapy completely eliminated tumors in most of the animals and significantly prolonged their survival.

“One of the examples where this particular technology could be useful is trying to control the growth of really fast-growing tumors,” says Ana Jaklenec, a principal investigator at MIT’s Koch Institute for Integrative Cancer Research. “The goal would be to gain some control over these tumors for patients that don't really have a lot of options, and this could either prolong their life or at least allow them to have a better quality of life during this period.”

Jaklenec is one of the senior authors of the new study, along with Angela Belcher, the James Mason Crafts Professor of Biological Engineering and Materials Science and Engineering and a member of the Koch Institute, and Robert Langer, an MIT Institute Professor and member of the Koch Institute. Maria Kanelli, a former MIT postdoc, is the lead author of the paper, which appears today in the journal ACS Nano.

Dual therapy

Patients with advanced tumors usually undergo a combination of treatments, including chemotherapy, surgery, and radiation. Phototherapy is a newer treatment that involves implanting or injecting particles that are heated with an external laser, raising their temperature enough to kill nearby tumor cells without damaging other tissue.

Current approaches to phototherapy in clinical trials make use of gold nanoparticles, which emit heat when exposed to near-infrared light.

The MIT team wanted to come up with a way to deliver phototherapy and chemotherapy together, which they thought could make the treatment process easier on the patient and might also have synergistic effects. They decided to use an inorganic material called molybdenum sulfide as the phototherapeutic agent. This material converts laser light to heat very efficiently, which means that low-powered lasers can be used.

To create a microparticle that could deliver both of these treatments, the researchers combined molybdenum disulfide nanosheets with either doxorubicin, a hydrophilic drug, or violacein, a hydrophobic drug. To make the particles, molybdenum disulfide and the chemotherapeutic are mixed with a polymer called polycaprolactone and then dried into a film that can be pressed into microparticles of different shapes and sizes.

For this study, the researchers created cubic particles with a width of 200 micrometers. Once injected into a tumor site, the particles remain there throughout the treatment. During each treatment cycle, an external near-infrared laser is used to heat up the particles. This laser can penetrate to a depth of a few millimeters to centimeters, with a local effect on the tissue.

“The advantage of this platform is that it can act on demand in a pulsatile manner,” Kanelli says. “You administer it once through an intratumoral injection, and then using an external laser source you can activate the platform, release the drug, and at the same time achieve thermal ablation of the tumor cells.”

To optimize the treatment protocol, the researchers used machine-learning algorithms to figure out the laser power, irradiation time, and concentration of the phototherapeutic agent that would lead to the best outcomes.

That led them to design a laser treatment cycle that lasts for about three minutes. During that time, the particles are heated to about 50 degrees Celsius, which is hot enough to kill tumor cells. Also at this temperature, the polymer matrix within the particles begins to melt, releasing some of the chemotherapy drug contained within the matrix.

“This machine-learning-optimized laser system really allows us to deploy low-dose, localized chemotherapy by leveraging the deep tissue penetration of near-infrared light for pulsatile, on-demand photothermal therapy. This synergistic effect results in low systemic toxicity compared to conventional chemotherapy regimens,” says Neelkanth Bardhan, a Break Through Cancer research scientist in the Belcher Lab, and second author of the paper.

Eliminating tumors

The researchers tested the microparticle treatment in mice that were injected with an aggressive type of cancer cells from triple-negative breast tumors. Once tumors formed, the researchers implanted about 25 microparticles per tumor, and then performed the laser treatment three times, with three days in between each treatment.

“This is a powerful demonstration of the usefulness of near-infrared-responsive material systems,” says Belcher, who, along with Bardhan, has previously worked on near-infrared imaging systems for diagnostic and treatment applications in ovarian cancer. “Controlling the drug release at timed intervals with light, after just one dose of particle injection, is a game changer for less painful treatment options and can lead to better patient compliance.”

In mice that received this treatment, the tumors were completely eradicated, and the mice lived much longer than those that were given either chemotherapy or phototherapy alone, or no treatment. Mice that underwent all three treatment cycles also fared much better than those that received just one laser treatment.

The polymer used to make the particles is biocompatible and has already been FDA-approved for medical devices. The researchers now hope to test the particles in larger animal models, with the goal of eventually evaluating them in clinical trials. They expect that this treatment could be useful for any type of solid tumor, including metastatic tumors.

The research was funded by the Bodossaki Foundation, the Onassis Foundation, a Mazumdar-Shaw International Oncology Fellowship, a National Cancer Institute Fellowship, and the Koch Institute Support (core) Grant from the National Cancer Institute.

In the classic cartoon “The Jetsons,” Rosie the robotic maid seamlessly switches from vacuuming the house to cooking dinner to taking out the trash. But in real life, training a general-purpose robot remains a major challenge.

Typically, engineers collect data that are specific to a certain robot and task, which they use to train the robot in a controlled environment. However, gathering these data is costly and time-consuming, and the robot will likely struggle to adapt to environments or tasks it hasn’t seen before.

To train better general-purpose robots, MIT researchers developed a versatile technique that combines a huge amount of heterogeneous data from many of sources into one system that can teach any robot a wide range of tasks.

Their method involves aligning data from varied domains, like simulations and real robots, and multiple modalities, including vision sensors and robotic arm position encoders, into a shared “language” that a generative AI model can process.

By combining such an enormous amount of data, this approach can be used to train a robot to perform a variety of tasks without the need to start training it from scratch each time.

This method could be faster and less expensive than traditional techniques because it requires far fewer task-specific data. In addition, it outperformed training from scratch by more than 20 percent in simulation and real-world experiments.

“In robotics, people often claim that we don’t have enough training data. But in my view, another big problem is that the data come from so many different domains, modalities, and robot hardware. Our work shows how you’d be able to train a robot with all of them put together,” says Lirui Wang, an electrical engineering and computer science (EECS) graduate student and lead author of a paper on this technique.

Wang’s co-authors include fellow EECS graduate student Jialiang Zhao; Xinlei Chen, a research scientist at Meta; and senior author Kaiming He, an associate professor in EECS and a member of the Computer Science and Artificial Intelligence Laboratory (CSAIL). The research will be presented at the Conference on Neural Information Processing Systems.

Inspired by LLMs

A robotic “policy” takes in sensor observations, like camera images or proprioceptive measurements that track the speed and position a robotic arm, and then tells a robot how and where to move.

Policies are typically trained using imitation learning, meaning a human demonstrates actions or teleoperates a robot to generate data, which are fed into an AI model that learns the policy. Because this method uses a small amount of task-specific data, robots often fail when their environment or task changes.

To develop a better approach, Wang and his collaborators drew inspiration from large language models like GPT-4.

These models are pretrained using an enormous amount of diverse language data and then fine-tuned by feeding them a small amount of task-specific data. Pretraining on so much data helps the models adapt to perform well on a variety of tasks.

“In the language domain, the data are all just sentences. In robotics, given all the heterogeneity in the data, if you want to pretrain in a similar manner, we need a different architecture,” he says.

Robotic data take many forms, from camera images to language instructions to depth maps. At the same time, each robot is mechanically unique, with a different number and orientation of arms, grippers, and sensors. Plus, the environments where data are collected vary widely.

The MIT researchers developed a new architecture called Heterogeneous Pretrained Transformers (HPT) that unifies data from these varied modalities and domains.

They put a machine-learning model known as a transformer into the middle of their architecture, which processes vision and proprioception inputs. A transformer is the same type of model that forms the backbone of large language models.

The researchers align data from vision and proprioception into the same type of input, called a token, which the transformer can process. Each input is represented with the same fixed number of tokens.

Then the transformer maps all inputs into one shared space, growing into a huge, pretrained model as it processes and learns from more data. The larger the transformer becomes, the better it will perform.

A user only needs to feed HPT a small amount of data on their robot’s design, setup, and the task they want it to perform. Then HPT transfers the knowledge the transformer grained during pretraining to learn the new task.

Enabling dexterous motions

One of the biggest challenges of developing HPT was building the massive dataset to pretrain the transformer, which included 52 datasets with more than 200,000 robot trajectories in four categories, including human demo videos and simulation.

The researchers also needed to develop an efficient way to turn raw proprioception signals from an array of sensors into data the transformer could handle.

“Proprioception is key to enable a lot of dexterous motions. Because the number of tokens is in our architecture always the same, we place the same importance on proprioception and vision,” Wang explains.

When they tested HPT, it improved robot performance by more than 20 percent on simulation and real-world tasks, compared with training from scratch each time. Even when the task was very different from the pretraining data, HPT still improved performance.

“This paper provides a novel approach to training a single policy across multiple robot embodiments. This enables training across diverse datasets, enabling robot learning methods to significantly scale up the size of datasets that they can train on. It also allows the model to quickly adapt to new robot embodiments, which is important as new robot designs are continuously being produced,” says David Held, associate professor at the Carnegie Mellon University Robotics Institute, who was not involved with this work.

In the future, the researchers want to study how data diversity could boost the performance of HPT. They also want to enhance HPT so it can process unlabeled data like GPT-4 and other large language models.

“Our dream is to have a universal robot brain that you could download and use for your robot without any training at all. While we are just in the early stages, we are going to keep pushing hard and hope scaling leads to a breakthrough in robotic policies, like it did with large language models,” he says.

This work was funded, in part, by the Amazon Greater Boston Tech Initiative and the Toyota Research Institute.

When you think about hands-free devices, you might picture Alexa and other voice-activated in-home assistants, Bluetooth earpieces, or asking Siri to make a phone call in your car. You might not imagine using your mouth to communicate with other devices like a computer or a phone remotely. 

Thinking outside the box, MIT Computer Science and Artificial Intelligence Laboratory (CSAIL) and Aarhus University researchers have now engineered “MouthIO,” a dental brace that can be fabricated with sensors and feedback components to capture in-mouth interactions and data. This interactive wearable could eventually assist dentists and other doctors with collecting health data and help motor-impaired individuals interact with a phone, computer, or fitness tracker using their mouths.

Resembling an electronic retainer, MouthIO is a see-through brace that fits the specifications of your upper or lower set of teeth from a scan. The researchers created a plugin for the modeling software Blender to help users tailor the device to fit a dental scan, where you can then 3D print your design in dental resin. This computer-aided design tool allows users to digitally customize a panel (called PCB housing) on the side to integrate electronic components like batteries, sensors (including detectors for temperature and acceleration, as well as tongue-touch sensors), and actuators (like vibration motors and LEDs for feedback). You can also place small electronics outside of the PCB housing on individual teeth.

The active mouth

“The mouth is a really interesting place for an interactive wearable and can open up many opportunities, but has remained largely unexplored due to its complexity,” says senior author Michael Wessely, a former CSAIL postdoc and senior author on a paper about MouthIO who is now an assistant professor at Aarhus University. “This compact, humid environment has elaborate geometries, making it hard to build a wearable interface to place inside. With MouthIO, though, we’ve developed a new kind of device that’s comfortable, safe, and almost invisible to others. Dentists and other doctors are eager about MouthIO for its potential to provide new health insights, tracking things like teeth grinding and potentially bacteria in your saliva.”

The excitement for MouthIO’s potential in health monitoring stems from initial experiments. The team found that their device could track bruxism (the habit of grinding teeth) by embedding an accelerometer within the brace to track jaw movements. When attached to the lower set of teeth, MouthIO detected when users grind and bite, with the data charted to show how often users did each.

Wessely and his colleagues’ customizable brace could one day help users with motor impairments, too. The team connected small touchpads to MouthIO, helping detect when a user’s tongue taps their teeth. These interactions could be sent via Bluetooth to scroll across a webpage, for example, allowing the tongue to act as a “third hand” to open up a new avenue for hands-free interaction.

"MouthIO is a great example how miniature electronics now allow us to integrate sensing into a broad range of everyday interactions,” says study co-author Stefanie Mueller, the TIBCO Career Development Associate Professor in the MIT departments of Electrical Engineering and Computer Science and Mechanical Engineering and leader of the HCI Engineering Group at CSAIL. “I'm especially excited about the potential to help improve accessibility and track potential health issues among users."

Molding and making MouthIO

To get a 3D model of your teeth, you can first create a physical impression and fill it with plaster. You can then scan your mold with a mobile app like Polycam and upload that to Blender. Using the researchers’ plugin within this program, you can clean up your dental scan to outline a precise brace design. Finally, you 3D print your digital creation in clear dental resin, where the electronic components can then be soldered on. Users can create a standard brace that covers their teeth, or opt for an “open-bite” design within their Blender plugin. The latter fits more like open-finger gloves, exposing the tips of your teeth, which helps users avoid lisping and talk naturally.

This “do it yourself” method costs roughly $15 to produce and takes two hours to be 3D-printed. MouthIO can also be fabricated with a more expensive, professional-level teeth scanner similar to what dentists and orthodontists use, which is faster and less labor-intensive.

Compared to its closed counterpart, which fully covers your teeth, the researchers view the open-bite design as a more comfortable option. The team preferred to use it for beverage monitoring experiments, where they fabricated a brace capable of alerting users when a drink was too hot. This iteration of MouthIO had a temperature sensor and a monitor embedded within the PCB housing that vibrated when a drink exceeded 65 degrees Celsius (or 149 degrees Fahrenheit). This could help individuals with mouth numbness better understand what they’re consuming.

In a user study, participants also preferred the open-bite version of MouthIO. “We found that our device could be suitable for everyday use in the future,” says study lead author and Aarhus University PhD student Yijing Jiang. “Since the tongue can touch the front teeth in our open-bite design, users don’t have a lisp. This made users feel more comfortable wearing the device during extended periods with breaks, similar to how people use retainers.”

The team’s initial findings indicate that MouthIO is a cost-effective, accessible, and customizable interface, and the team is working on a more long-term study to evaluate its viability further. They’re looking to improve its design, including experimenting with more flexible materials, and placing it in other parts of the mouth, like the cheek and the palate. Among these ideas, the researchers have already prototyped two new designs for MouthIO: a single-sided brace for even higher comfort when wearing MouthIO while also being fully invisible to others, and another fully capable of wireless charging and communication.

Jiang, Mueller, and Wessely’s co-authors include PhD student Julia Kleinau, master’s student Till Max Eckroth, and associate professor Eve Hoggan, all of Aarhus University. Their work was supported by a Novo Nordisk Foundation grant and was presented at ACM’s Symposium on User Interface Software and Technology.

Hospice care aims to provide a health care alternative for people nearing the end of life by sparing them unwanted medical procedures and focusing on the patient’s comfort. A new study co-authored by MIT scholars shows hospice also has a clear fiscal benefit: It generates substantial savings for the U.S. Medicare system.

The study examines the growth of for-profit hospice providers, who receive reimbursements from Medicare, and evaluates the cost of caring for patients with Alzheimer’s disease and related dementias (ADRD). The research finds that for patients using for-profit hospice providers, there is about a $29,000 savings to Medicare over the first five years after someone is diagnosed with ADRD.

“Hospice is saving Medicare a lot of money,” says Jonathan Gruber, an MIT health care economist and co-author of a paper detailing the study’s findings. “Those are big numbers.”

In recent decades, hospice care has grown substantially. That growth has been accompanied by concerns that for-profit hospice organizations, in particular, might be overly aggressive in pursuing patients. There have also been instances of fraud by organizations in the field. And yet, the study shows that the overall dynamics of hospice are the intended ones: People are indeed receiving palliative-type care, based around comfort rather than elaborate medical procedures, at less cost.

“What we found is that hospice basically operates as advertised,” adds Gruber, the Ford Professor of Economics at MIT. “It does not extend lives on aggregate, and it does save money.”

The paper, “Dying or Lying? For-Profit Hospices and End of Life Care,” appears in the American Economic Review. The co-authors are Gruber, who is also head of MIT’s Department of Economics; David Howard, a professor at the Rollins School of Public Health at Emory University; Jetson Leder-Luis PhD ’20, an assistant professor at Boston University; and Theodore Caputi, a doctoral student in MIT’s Department of Economics.

Charting what more hospice access means

Hospice care in the U.S. dates to at least the 1970s. Patients opt out of their existing medical network and receive nursing care where they live, either at home or in care facilities. That care is oriented around reducing suffering and pain, rather than attempting to eliminate underlying causes. Generally, hospice patients are expected to have six months or less to live. Most Medicare funding goes to private contractors supplying medical care, and in the 1980s the federal government started using Medicare to reimburse the medical expenses from hospice as well.

While the number of nonprofit hospice providers in the U.S. has remained fairly consistent, the number of for-profit hospice organizations grew fivefold between 2000 and 2019. Medicare payments for hospice care are now about $20 billion annually, up from $2.5 billion in 1999. People diagnosed with ADRD now make up 38 percent of hospice patients.

Still, Gruber considers the topic of hospice care relatively under-covered by analysts. To conduct the study, the team examined over 10 million patients from 1999 through 2019. The researchers used the growth of for-profit hospice providers to compare the effects of being enrolled in non-profit hospice care, for-profit hospice care, or staying in the larger medical system.

That means the scholars were not only evaluating hospice patients; by evaluating the larger population in a given area where and when for-profit hospice firms opened their doors, they could see what difference greater access to hospice care made. For instance, having a new for-profit hospice open locally is associated with a roughly 2 percentage point increase in for-profit hospice admissions in following years.

“We’re able to use this methodology to [analyze] if these patients would otherwise have not gone to hospice or would have gone to a nonprofit hospice,” Gruber says.

The method also allows the scholars to estimate the substantial cost savings. And it shows that enrolling in hospice increased the five-year post-diagnosis mortality rate of ADRD patients by 8.6 percentage points, from a baseline of 66.6 percent. Entering into hospice care — which is a reversible decision — means foregoing life-extending surgeries, for instance, if people believe such procedures are no longer desirable for them.

Rethinking the cap

By providing care without more expensive medical procedures, it is understandable that hospice reduces overall medical costs. Still, given that Medicare reimburses hospice organizations, one ongoing policy concern is that hospice providers might aggressively recruit a larger percentage of patients who end up living longer than six additional months. In this way hospice providers might unduly boost their revenues and put more pressure on the Medicare budget.

To counteract this, Medicare rules include a roughly $29,205 cap on per-patient reimbursements, as of 2019. Most patients die relatively soon after entering hospice care; some will outlive the six-month expectation significantly. But hospice organizations cannot exceed that average.

However, the study also suggests the cap is a suboptimal approach. In 2018, 15.5 percent of hospice patients were being discharged from hospice care while still alive, due to the cap limiting hospice capacity. As the paper notes, “patients in hospices facing cap pressure are more likely to be discharged from hospice alive and experience higher mortality rates.”

As Gruber notes, the spending cap is partly a fraud-fighting tool. And yet the cap clearly has other, unintended consquences on patients and their medical choices, crowding some out of the hospice system.

“The cap may be throwing the baby out with the bathwater.” Gruber says. “The government has more focused tools to fight fraud. Using the cap for that is a blunt instrument.”

As long as people are informed about hospice and the medical trajectory it puts them on, then, hospice care appears to be providing a valued service at less expense than other approaches to end-of-life care.

“The holy grail in health care is things that improve quality and save money,” Gruber says. “And with hospice, there are surveys saying people like it. And it certainly saves money, and there’s no evidence it’s doing harm [to patients]. We talk about how we struggle to deal with health care costs in this country, so this seems like what we want.”

The research was supported in part by the National Institute on Aging of the National Institutes of Health. 

A team led by researchers at MIT has discovered that a distant interstellar cloud contains an abundance of pyrene, a type of large, carbon-containing molecule known as a polycyclic aromatic hydrocarbon (PAH).

The discovery of pyrene in this far-off cloud, which is similar to the collection of dust and gas that eventually became our own solar system, suggests that pyrene may have been the source of much of the carbon in our solar system. That hypothesis is also supported by a recent finding that samples returned from the near-Earth asteroid Ryugu contain large quantities of pyrene.

“One of the big questions in star and planet formation is: How much of the chemical inventory from that early molecular cloud is inherited and forms the base components of the solar system? What we’re looking at is the start and the end, and they’re showing the same thing. That’s pretty strong evidence that this material from the early molecular cloud finds its way into the ice, dust, and rocky bodies that make up our solar system,” says Brett McGuire, an assistant professor of chemistry at MIT.

Due to its symmetry, pyrene itself is invisible to the radio astronomy techniques that have been used to detect about 95 percent of molecules in space. Instead, the researchers detected an isomer of cyanopyrene, a version of pyrene that has reacted with cyanide to break its symmetry. The molecule was detected in a distant cloud known as TMC-1, using the 100-meter Green Bank Telescope (GBT), a radio telescope at the Green Bank Observatory in West Virginia.

McGuire and Ilsa Cooke, an assistant professor of chemistry at the University of British Colombia, are the senior authors of a paper describing the findings, which appears today in Science. Gabi Wenzel, an MIT postdoc in McGuire’s group, is the lead author of the study.

Carbon in space

PAHs, which contain rings of carbon atoms fused together, are believed to store 10 to 25 percent of the carbon that exists in space. More than 40 years ago, scientists using infrared telescopes began detecting features that are thought to belong to vibrational modes of PAHs in space, but this technique couldn’t reveal exactly which types of PAHs were out there.

“Since the PAH hypothesis was developed in the 1980s, many people have accepted that PAHs are in space, and they have been found in meteorites, comets, and asteroid samples, but we can’t really use infrared spectroscopy to unambiguously identify individual PAHs in space,” Wenzel says.

In 2018, a team led by McGuire reported the discovery of benzonitrile — a six-carbon ring attached to a nitrile (carbon-nitrogen) group — in TMC-1. To make this discovery, they used the GBT, which can detect molecules in space by their rotational spectra — distinctive patterns of light that molecules give off as they tumble through space. In 2021, his team detected the first individual PAHs in space: two isomers of cyanonaphthalene, which consists of two rings fused together, with a nitrile group attached to one ring.

On Earth, PAHs commonly occur as byproducts of burning fossil fuels, and they’re also found in char marks on grilled food. Their discovery in TMC-1, which is only about 10 kelvins, suggested that it may also be possible for them to form at very low temperatures.

The fact that PAHs have also been found in meteorites, asteroids, and comets has led many scientists to hypothesize that PAHs are the source of much of the carbon that formed our own solar system. In 2023, researchers in Japan found large quantities of pyrene in samples returned from the asteroid Ryugu during the Hayabusa2 mission, along with smaller PAHs including naphthalene.

That discovery motivated McGuire and his colleagues to look for pyrene in TMC-1. Pyrene, which contains four rings, is larger than any of the other PAHs that have been detected in space. In fact, it’s the third-largest molecule identified in space, and the largest ever detected using radio astronomy.

Before looking for these molecules in space, the researchers first had to synthesize cyanopyrene in the laboratory. The cyano or nitrile group is necessary for the molecule to emit a signal that a radio telescope can detect. The synthesis was performed by MIT postdoc Shuo Zhang in the group of Alison Wendlandt, an MIT associate professor of chemistry.

Then, the researchers analyzed the signals that the molecules emit in the laboratory, which are exactly the same as the signals that they emit in space.

Using the GBT, the researchers found these signatures throughout TMC-1. They also found that cyanopyrene accounts for about 0.1 percent of all the carbon found in the cloud, which sounds small but is significant when one considers the thousands of different types of carbon-containing molecules that exist in space, McGuire says.

“While 0.1 percent doesn’t sound like a large number, most carbon is trapped in carbon monoxide (CO), the second-most abundant molecule in the universe besides molecular hydrogen. If we set CO aside, one in every few hundred or so remaining carbon atoms is in pyrene. Imagine the thousands of different molecules that are out there, nearly all of them with many different carbon atoms in them, and one in a few hundred is in pyrene,” he says. “That is an absolutely massive abundance. An almost unbelievable sink of carbon. It’s an interstellar island of stability.”

Ewine van Dishoeck, a professor of molecular astrophysics at Leiden Observatory in the Netherlands, called the discovery “unexpected and exciting.”

“It builds on their earlier discoveries of smaller aromatic molecules, but to make the jump now to the pyrene family is huge. Not only does it demonstrate that a significant fraction of carbon is locked up in these molecules, but it also points to different formation routes of aromatics than have been considered so far,” says van Dishoeck, who was not involved in the research.

An abundance of pyrene

Interstellar clouds like TMC-1 may eventually give rise to stars, as clumps of dust and gas coalesce into larger bodies and begin to heat up. Planets, asteroids, and comets arise from some of the gas and dust that surround young stars. Scientists can’t look back in time at the interstellar cloud that gave rise to our own solar system, but the discovery of pyrene in TMC-1, along with the presence of large amounts of pyrene in the asteroid Ryugu, suggests that pyrene may have been the source of much of the carbon in our own solar system.

“We now have, I would venture to say, the strongest evidence ever of this direct molecular inheritance from the cold cloud all the way through to the actual rocks in the solar system,” McGuire says.

The researchers now plan to look for even larger PAH molecules in TMC-1. They also hope to investigate the question of whether the pyrene found in TMC-1 was formed within the cold cloud or whether it arrived from elsewhere in the universe, possibly from the high-energy combustion processes that surround dying stars.

The research was funded in part by a Beckman Foundation Young Investigator Award, the Schmidt Futures, the U.S. National Science Foundation, the Natural Sciences and Engineering Research Council of Canada, the Goddard Center for Astrobiology, and the NASA Planetary Science Division Internal Scientist Funding Program.

For many decades, fusion has been touted as the ultimate source of abundant, clean electricity. Now, as the world faces the need to reduce carbon emissions to prevent catastrophic climate change, making commercial fusion power a reality takes on new importance. In a power system dominated by low-carbon variable renewable energy sources (VREs) such as solar and wind, “firm” electricity sources are needed to kick in whenever demand exceeds supply — for example, when the sun isn’t shining or the wind isn’t blowing and energy storage systems aren’t up to the task. What is the potential role and value of fusion power plants (FPPs) in such a future electric power system — a system that is not only free of carbon emissions but also capable of meeting the dramatically increased global electricity demand expected in the coming decades?

Working together for a year-and-a-half, investigators in the MIT Energy Initiative (MITEI) and the MIT Plasma Science and Fusion Center (PSFC) have been collaborating to answer that question. They found that — depending on its future cost and performance — fusion has the potential to be critically important to decarbonization. Under some conditions, the availability of FPPs could reduce the global cost of decarbonizing by trillions of dollars. More than 25 experts together examined the factors that will impact the deployment of FPPs, including costs, climate policy, operating characteristics, and other factors. They present their findings in a new report funded through MITEI and entitled “The Role of Fusion Energy in a Decarbonized Electricity System.”

“Right now, there is great interest in fusion energy in many quarters — from the private sector to government to the general public,” says the study’s principal investigator (PI) Robert C. Armstrong, MITEI’s former director and the Chevron Professor of Chemical Engineering, Emeritus. “In undertaking this study, our goal was to provide a balanced, fact-based, analysis-driven guide to help us all understand the prospects for fusion going forward.” Accordingly, the study takes a multidisciplinary approach that combines economic modeling, electric grid modeling, techno-economic analysis, and more to examine important factors that are likely to shape the future deployment and utilization of fusion energy. The investigators from MITEI provided the energy systems modeling capability, while the PSFC participants provided the fusion expertise.

Fusion technologies may be a decade away from commercial deployment, so the detailed technology and costs of future commercial FPPs are not known at this point. As a result, the MIT research team focused on determining what cost levels fusion plants must reach by 2050 to achieve strong market penetration and make a significant contribution to the decarbonization of global electricity supply in the latter half of the century.

The value of having FPPs available on an electric grid will depend on what other options are available, so to perform their analyses, the researchers needed estimates of the future cost and performance of those options, including conventional fossil fuel generators, nuclear fission power plants, VRE generators, and energy storage technologies, as well as electricity demand for specific regions of the world. To find the most reliable data, they searched the published literature as well as results of previous MITEI and PSFC analyses.

Overall, the analyses showed that — while the technology demands of harnessing fusion energy are formidable — so are the potential economic and environmental payoffs of adding this firm, low-carbon technology to the world’s portfolio of energy options.

Perhaps the most remarkable finding is the “societal value” of having commercial FPPs available. “Limiting warming to 1.5 degrees C requires that the world invest in wind, solar, storage, grid infrastructure, and everything else needed to decarbonize the electric power system,” explains Randall Field, executive director of the fusion study and MITEI’s director of research. “The cost of that task can be far lower when FPPs are available as a source of clean, firm electricity.” And the benefit varies depending on the cost of the FPPs. For example, assuming that the cost of building a FPP is $8,000 per kilowatt (kW) in 2050 and falls to $4,300/kW in 2100, the global cost of decarbonizing electric power drops by $3.6 trillion. If the cost of a FPP is $5,600/kW in 2050 and falls to $3,000/kW in 2100, the savings from having the fusion plants available would be $8.7 trillion. (Those calculations are based on differences in global gross domestic product and assume a discount rate of 6 percent. The undiscounted value is about 20 times larger.)

The goal of other analyses was to determine the scale of deployment worldwide at selected FPP costs. Again, the results are striking. For a deep decarbonization scenario, the total global share of electricity generation from fusion in 2100 ranges from less than 10 percent if the cost of fusion is high to more than 50 percent if the cost of fusion is low.

Other analyses showed that the scale and timing of fusion deployment vary in different parts of the world. Early deployment of fusion can be expected in wealthy nations such as European countries and the United States that have the most aggressive decarbonization policies. But certain other locations — for example, India and the continent of Africa — will have great growth in fusion deployment in the second half of the century due to a large increase in demand for electricity during that time. “In the U.S. and Europe, the amount of demand growth will be low, so it’ll be a matter of switching away from dirty fuels to fusion,” explains Sergey Paltsev, deputy director of the MIT Center for Sustainability Science and Strategy and a senior research scientist at MITEI. “But in India and Africa, for example, the tremendous growth in overall electricity demand will be met with significant amounts of fusion along with other low-carbon generation resources in the later part of the century.”

A set of analyses focusing on nine subregions of the United States showed that the availability and cost of other low-carbon technologies, as well as how tightly carbon emissions are constrained, have a major impact on how FPPs would be deployed and used. In a decarbonized world, FPPs will have the highest penetration in locations with poor diversity, capacity, and quality of renewable resources, and limits on carbon emissions will have a big impact. For example, the Atlantic and Southeast subregions have low renewable resources. In those subregions, wind can produce only a small fraction of the electricity needed, even with maximum onshore wind buildout. Thus, fusion is needed in those subregions, even when carbon constraints are relatively lenient, and any available FPPs would be running much of the time. In contrast, the Central subregion of the United States has excellent renewable resources, especially wind. Thus, fusion competes in the Central subregion only when limits on carbon emissions are very strict, and FPPs will typically be operated only when the renewables can’t meet demand.

An analysis of the power system that serves the New England states provided remarkably detailed results. Using a modeling tool developed at MITEI, the fusion team explored the impact of using different assumptions about not just cost and emissions limits but even such details as potential land-use constraints affecting the use of specific VREs. This approach enabled them to calculate the FPP cost at which fusion units begin to be installed. They were also able to investigate how that “threshold” cost changed with changes in the cap on carbon emissions. The method can even show at what price FPPs begin to replace other specific generating sources. In one set of runs, they determined the cost at which FPPs would begin to displace floating platform offshore wind and rooftop solar.

“This study is an important contribution to fusion commercialization because it provides economic targets for the use of fusion in the electricity markets,” notes Dennis G. Whyte, co-PI of the fusion study, former director of the PSFC, and the Hitachi America Professor of Engineering in the Department of Nuclear Science and Engineering. “It better quantifies the technical design challenges for fusion developers with respect to pricing, availability, and flexibility to meet changing demand in the future.”

The researchers stress that while fission power plants are included in the analyses, they did not perform a “head-to-head” comparison between fission and fusion, because there are too many unknowns. Fusion and nuclear fission are both firm, low-carbon electricity-generating technologies; but unlike fission, fusion doesn’t use fissile materials as fuels, and it doesn’t generate long-lived nuclear fuel waste that must be managed. As a result, the regulatory requirements for FPPs will be very different from the regulations for today’s fission power plants — but precisely how they will differ is unclear. Likewise, the future public perception and social acceptance of each of these technologies cannot be projected, but could have a major influence on what generation technologies are used to meet future demand.

The results of the study convey several messages about the future of fusion. For example, it’s clear that regulation can be a potentially large cost driver. This should motivate fusion companies to minimize their regulatory and environmental footprint with respect to fuels and activated materials. It should also encourage governments to adopt appropriate and effective regulatory policies to maximize their ability to use fusion energy in achieving their decarbonization goals. And for companies developing fusion technologies, the study’s message is clearly stated in the report: “If the cost and performance targets identified in this report can be achieved, our analysis shows that fusion energy can play a major role in meeting future electricity needs and achieving global net-zero carbon goals.”

Researchers from the Critical Analytics for Manufacturing Personalized-Medicine (CAMP) interdisciplinary research group at the Singapore-MIT Alliance for Research and Technology (SMART), MIT’s research enterprise in Singapore, alongside collaborators from the National University of Singapore Tissue Engineering Programme, have developed a novel method to enhance the ability of mesenchymal stromal cells (MSCs) to generate cartilage tissue by adding ascorbic acid during MSC expansion. The research also discovered that micro-magnetic resonance relaxometry (µMRR), a novel process analytical tool developed by SMART CAMP, can be used as a rapid, label-free process-monitoring tool for the quality expansion of MSCs.

Articular cartilage, a connective tissue that protects the bone ends in joints, can degenerate due to injury, age, or arthritis, leading to significant joint pain and disability. Especially in countries — such as Singapore — that have an active, aging population, articular cartilage degeneration is a growing ailment that affects an increasing number of people. Autologous chondrocyte implantation is currently the only Food and Drug Administration-approved cell-based therapy for articular cartilage injuries, but it is costly, time-intensive, and requires multiple treatments. MSCs are an attractive and promising alternative as they have shown good safety profiles for transplantation. However, clinical use of MSCs is limited due to inconsistent treatment outcomes arising from factors such as donor-to-donor variability, variation among cells during cell expansion, and non-standardized MSC manufacturing protocols.

The heterogeneity of MSCs can lead to variations in their biological behavior and treatment outcomes. While large-scale MSC expansions are required to obtain a therapeutically relevant number of cells for implantation, this process can introduce cell heterogeneity. Therefore, improved processes are essential to reduce cell heterogeneity while increasing donor cell numbers with improved chondrogenic potential — the ability of MSCs to differentiate into cartilage cells to repair cartilage tissue — to pave the way for more effective and consistent MSC-based therapies.

In a paper titled “Metabolic modulation to improve MSC expansion and therapeutic potential for articular cartilage repair,” published in the scientific journal Stem Cell Research and Therapy, CAMP researchers detailed their development of a priming strategy to enhance the expansion of quality MSCs by modifying the way cells utilize energy. The research findings have shown a positive correlation between chondrogenic potential and oxidative phosphorylation (OXPHOS), a process that harnesses the reduction of oxygen to create adenosine triphosphate — a source of energy that drives and supports many processes in living cells. This suggests that manipulating MSC metabolism is a promising strategy for enhancing chondrogenic potential.

Using novel PATs developed by CAMP, the researchers explored the potential of metabolic modulation in both short- and long-term harvesting and reseeding of cells. To enhance their chondrogenic potential, they varied the nutrient composition, including glucose, pyruvate, glutamine, and ascorbic acid (AA). As AA is reported to support OXPHOS and its positive impact on chondrogenic potential during differentiation — a process in which immature cells become mature cells with specific functions — the researchers further investigated its effects during MSC expansion.

The addition of AA to cell cultures for one passage during MSC expansion and prior to initiation of differentiation was found to improve chondrogenic differentiation, which is a critical quality attribute (CQA) for better articular cartilage repair. Longer-term AA treatment led to a more than 300-fold increase in the yield of MSCs with enhanced chondrogenic potential, and reduced cell heterogeneity and cell senescence — a process by which a cell ages and permanently stops dividing but does not die — when compared to untreated cells. AA-treated MSCs with improved chondrogenic potential showed a robust shift in metabolic profile to OXPHOS. This metabolic change correlated with μMRR measurements, which helps identify novel CQAs that could be implemented in MSC manufacturing for articular cartilage repair.

The research also demonstrates the potential of the process analytical tool developed by CAMP, micromagnetic resonance relaxometry (μMRR) — a miniature benchtop device that employs magnetic resonance imaging (MRI) imaging on a microscopic scale — as a process-monitoring tool for the expansion of MSCs with AA supplementation. Originally used as a label-free malaria diagnosis method due to the presence of paramagnetic hemozoin particles, μMRR was used in the research to detect senescence in MSCs. This rapid, label-free method requires only a small number of cells for evaluation, which allows for MSC therapy manufacturing in closed systems — a system for protecting pharmaceutical products by reducing contamination risks from the external environment — while enabling intermittent monitoring of a limited lot size per production.

“Donor-to-donor variation, intrapopulation heterogeneity, and cellular senescence have impeded the success of MSCs as a standard of care therapy for articular cartilage repair. Our research showed that AA supplementation during MSC expansion can overcome these bottlenecks and enhance MSC chondrogenic potential,” says Ching Ann Tee, senior postdoc at SMART CAMP and first author of the paper. “By controlling metabolic conditions such as AA supplementation, coupled with CAMP’s process analytical tools such as µMRR, the yield and quality of cell therapy products could be significantly increased. This breakthrough could help make MSC therapy a more effective and viable treatment option and provide standards for improving the manufacturing pipeline.”

“This approach of utilizing metabolic modulation to improve MSC chondrogenic potential could be adapted into similar concepts for other therapeutic indications, such as osteogenic potential for bone repair or other types of stem cells. Implementing our findings in MSC manufacturing settings could be a significant step forward for patients with osteoarthritis and other joint diseases, as we can efficiently produce large quantities of high-quality MSCs with consistent functionality and enable the treatment of more patients,” adds Professor Laurie A. Boyer, principal investigator at SMART CAMP, professor of biology and biological engineering at MIT, and corresponding author of the paper.

The research is conducted by SMART and supported by the National Research Foundation Singapore under its Campus for Research Excellence and Technological Enterprise program.

In a first for both universities, MIT undergraduates are engaged in research projects at the Universidad del Valle de Guatemala (UVG), while MIT scholars are collaborating with UVG undergraduates on in-depth field studies in Guatemala.

These pilot projects are part of a larger enterprise, called ASPIRE (Achieving Sustainable Partnerships for Innovation, Research, and Entrepreneurship). Funded by the U.S. Agency for International Development, this five-year, $15-million initiative brings together MIT, UVG, and the Guatemalan Exporters Association to promote sustainable solutions to local development challenges.

“This research is yielding insights into our understanding of how to design with and for marginalized people, specifically Indigenous people,” says Elizabeth Hoffecker, co-principal investigator of ASPIRE at MIT and director of the MIT Local Innovation Group.

The students’ work is bearing fruit in the form of publications and new products — directly advancing ASPIRE’s goals to create an innovation ecosystem in Guatemala that can be replicated elsewhere in Central and Latin America.

For the students, the project offers rewards both tangible and inspirational.

“My experience allowed me to find my interest in local innovation and entrepreneurship,” says Ximena Sarmiento García, a fifth-year undergraduate at UVG majoring in anthropology. Supervised by Hoffecker, Sarmiento García says, “I learned how to inform myself, investigate, and find solutions — to become a researcher.”

Sandra Youssef, a rising junior in mechanical engineering at MIT, collaborated with UVG researchers and Indigenous farmers to design a mobile cart to improve the harvest yield of snow peas. “It was perfect for me,” she says. “My goal was to use creative, new technologies and science to make a dent in difficult problems.”

Remote and effective

Kendra Leith, co-principal investigator of ASPIRE, and associate director for research at MIT D-Lab, shaped the MIT-based undergraduate research opportunities (UROPs) in concert with UVG colleagues. “Although MIT students aren’t currently permitted to travel to Guatemala, I wanted them to have an opportunity to apply their experience and knowledge to address real-world challenges,” says Leith. “The Covid pandemic prepared them and their counterparts at UVG for effective remote collaboration — the UROPs completed remarkably productive research projects over Zoom and met our goals for them.”

MIT students participated in some of UVG’s most ambitious ASPIRE research. For instance, Sydney Baller, a rising sophomore in mechanical engineering, joined a team of Indigenous farmers and UVG mechanical engineers investigating the manufacturing process and potential markets for essential oils extracted from thyme, rosemary, and chamomile plants.

“Indigenous people have thousands of years working with plant extracts and ancient remedies,” says Baller. “There is promising history there that would be important to follow up with more modern research.”

Sandra Youssef used computer-aided design and manufacturing to realize a design created in a hackathon by snow pea farmers. “Our cart had to hold 495 pounds of snow peas without collapsing or overturning, navigate narrow paths on hills, and be simple and inexpensive to assemble,” she says. The snow pea producers have tested two of Youssef’s designs, built by a team at UVG led by Rony Herrarte, a faculty member in the department of mechanical engineering.

From waste to filter

Two MIT undergraduates joined one of UVG’s long-standing projects: addressing pollution in Guatemala’s water. The research seeks to use chitosan molecules, extracted from shrimp shells, for bioremediation of heavy metals and other water contaminants. These shells are available in abundance, left as waste by the country’s shrimp industry.

Sophomores Ariana Hodlewsky, majoring in chemical engineering, and Paolo Mangiafico, majoring in brain and cognitive sciences, signed on to work with principal investigator and chemistry department instructor Allan Vásquez (UVG) on filtration systems utilizing chitosan.

“The team wants to find a cost-effective product rural communities, most at risk from polluted water, can use in homes or in town water systems,” says Mangiafico. “So we have been investigating different technologies for water filtration, and analyzing the Guatemalan and U.S. markets to understand the regulations and opportunities that might affect introduction of a chitosan-based product.”

“Our research into how different communities use water and into potential consumers and pitfalls sets the scene for prototypes UVG wants to produce,” says Hodlewsky.

Lourdes Figueroa, UVG ASPIRE project manager for technology transfer, found their assistance invaluable.

“Paolo and Ariana brought the MIT culture and mindset to the project,” she says. “They wanted to understand not only how the technology works, but the best ways of getting the technology out of the lab to make it useful.”

This was an “Aha!” moment, says Figueroa. “The MIT students made a major contribution to both the engineering and marketing sides by emphasizing that you have to think about how to guarantee the market acceptance of the technology while it is still under development.”

Innovation ecosystems

UVG’s three campuses have served as incubators for problem-solving innovation and entrepreneurship, in many cases driven by students from Indigenous communities and families. In 2022, Elizabeth Hoffecker, with eight UVG anthropology majors, set out to identify the most vibrant examples of these collaborative initiatives, which ASPIRE seeks to promote and replicate.

Hoffecker’s “innovation ecosystem diagnostic” revealed a cluster of activity centered on UVG’s Altiplano campus in the central highlands, which serves Mayan communities. Hoffecker and two of the anthropology students focused on four examples for a series of case studies, which they are currently preparing for submission to a peer-reviewed journal.

“The caliber of their work was so good that it became clear to me that we could collaborate on a paper,” says Hoffecker. “It was my first time publishing with undergraduates.”

The researchers’ cases included novel production of traditional thread, and creation of a 3D phytoplankton kit that is being used to educate community members about water pollution in Lake Atitlán, a tourist destination that drives the local economy but is increasingly being affected by toxic algae blooms. Hoffecker singles out a project by Indigenous undergraduates who developed play-based teaching tools for introducing basic mathematical concepts.

“These connect to local Mayan ways of understanding and offer a novel, hands-on way to strengthen the math teaching skills of local primary school teachers in Indigenous communities,” says Hoffecker. “They created something that addresses a very immediate need in the community — lack of training.

Both of Hoffecker’s undergraduate collaborators are writing theses inspired by these case studies.

“My time with Elizabeth allowed me to learn how to conduct research from scratch, ask for help, find solutions, and trust myself,” says Sarmiento García. She finds the ASPIRE approach profoundly appealing. “It is not only ethical, but also deeply committed to applying results to the real lives of the people involved.”

“This experience has been incredibly positive, validating my own ability to generate knowledge through research, rather than relying only on established authors to back up my arguments,” says Camila del Cid, a fifth-year anthropology student. “This was empowering, especially as a Latin American researcher, because it emphasized that my perspective and contributions are important.”

Hoffecker says this pilot run with UVG undergrads produced “high-quality research that can inform evidence-based decision-making on development issues of top regional priority” — a key goal for ASPIRE. Hoffecker plans to “develop a pathway that other UVG students can follow to conduct similar research.”

MIT undergraduate research will continue. “Our students’ activities have been very valuable in Guatemala, so much so that the snow pea, chitosan, and essential oils teams would like to continue working with our students this year,” says Leith.  She anticipates a new round of MIT UROPs for next summer.

Youssef, for one, is eager to get to work on refining the snow pea cart. “I like the idea of working outside my comfort zone, thinking about things that seem unsolvable and coming up with a solution to fix some aspect of the problem,” she says.

Many black holes detected to date appear to be part of a pair. These binary systems comprise a black hole and a secondary object — such as a star, a much denser neutron star, or another black hole — that spiral around each other, drawn together by the black hole’s gravity to form a tight orbital pair.

Now a surprising discovery is expanding the picture of black holes, the objects they can host, and the way they form.

In a study appearing today in Nature, physicists at MIT and Caltech report that they have observed a “black hole triple” for the first time. The new system holds a central black hole in the act of consuming a small star that’s spiraling in very close to the black hole, every 6.5 days — a configuration similar to most binary systems. But surprisingly, a second star appears to also be circling the black hole, though at a much greater distance. The physicists estimate this far-off companion is orbiting the black hole every 70,000 years.

That the black hole seems to have a gravitational hold on an object so far away is raising questions about the origins of the black hole itself. Black holes are thought to form from the violent explosion of a dying star — a process known as a supernova, by which a star releases a huge amount of energy and light in a final burst before collapsing into an invisible black hole.

The team’s discovery, however, suggests that if the newly-observed black hole resulted from a typical supernova, the energy it would have released before it collapsed would have kicked away any loosely bound objects in its outskirts. The second, outer star, then, shouldn’t still be hanging around.

Instead, the team suspects the black hole formed through a more gentle process of “direct collapse,” in which a star simply caves in on itself, forming a black hole without a last dramatic flash. Such a gentle origin would hardly disturb any loosely bound, faraway objects.

Because the new triple system includes a very far-off star, this suggests the system’s black hole was born through a gentler, direct collapse. And while astronomers have observed more violent supernovae for centuries, the team says the new triple system could be the first evidence of a black hole that formed from this more gentle process.

“We think most black holes form from violent explosions of stars, but this discovery helps call that into question,” says study author Kevin Burdge, a Pappalardo Fellow in the MIT Department of Physics. “This system is super exciting for black hole evolution, and it also raises questions of whether there are more triples out there.”

The study’s co-authors at MIT are Erin Kara, Claude Canizares, Deepto Chakrabarty, Anna Frebel, Sarah Millholland, Saul Rappaport, Rob Simcoe, and Andrew Vanderburg, along with Kareem El-Badry at Caltech.

Tandem motion

The discovery of the black hole triple came about almost by chance. The physicists found it while looking through Aladin Lite, a repository of astronomical observations, aggregated from telescopes in space and all around the world. Astronomers can use the online tool to search for images of the same part of the sky, taken by different telescopes that are tuned to various wavelengths of energy and light.

The team had been looking within the Milky Way galaxy for signs of new black holes. Out of curiosity, Burdge reviewed an image of V404 Cygni — a black hole about 8,000 light years from Earth that was one of the very first objects ever to be confirmed as a black hole, in 1992. Since then, V404 Cygni has become one of the most well-studied black holes, and has been documented in over 1,300 scientific papers. However, none of those studies reported what Burdge and his colleagues observed. 

As he looked at optical images of V404 Cygni, Burdge saw what appeared to be two blobs of light, surprisingly close to each other. The first blob was what others determined to be the black hole and an inner, closely orbiting star. The star is so close that it is shedding some of its material onto the black hole, and giving off the light that Burdge could see. The second blob of light, however, was something that scientists did not investigate closely, until now. That second light, Burdge determined, was most likely coming from a very far-off star.

“The fact that we can see two separate stars over this much distance actually means that the stars have to be really very far apart,” says Burdge, who calculated that the outer star is 3,500 astronomical units (AU) away from the black hole (1 AU is the distance between the Earth and sun). In other words, the outer star is 3,500 times father away from the black hole as the Earth is from the sun. This is also equal to 100 times the distance between Pluto and the sun.

The question that then came to mind was whether the outer star was linked to the black hole and its inner star. To answer this, the researchers looked to Gaia, a satellite that has precisely tracked the motions of all the stars in the galaxy since 2014. The team analyzed the motions of the inner and outer stars over the last 10 years of Gaia data and found that the stars moved exactly in tandem, compared to other neighboring stars. They calculated that the odds of this kind of tandem motion are about one in 10 million.

“It’s almost certainly not a coincidence or accident,” Burdge says. “We’re seeing two stars that are following each other because they’re attached by this weak string of gravity. So this has to be a triple system.”

Pulling strings

How, then, could the system have formed? If the black hole arose from a typical supernova, the violent explosion would have kicked away the outer star long ago.

“Imagine you’re pulling a kite, and instead of a strong string, you’re pulling with a spider web,” Burdge says. “If you tugged too hard, the web would break and you’d lose the kite. Gravity is like this barely bound string that’s really weak, and if you do anything dramatic to the inner binary, you’re going to lose the outer star.”

To really test this idea, however, Burdge carried out simulations to see how such a triple system could have evolved and retained the outer star.

At the start of each simulation, he introduced three stars (the third being the black hole, before it became a black hole). He then ran tens of thousands of simulations, each one with a slightly different scenario for how the third star could have become a black hole, and subsequently affected the motions of the other two stars. For instance, he simulated a supernova, varying the amount and direction of energy that it gave off. He also simulated scenarios of direct collapse, in which the third star simply caved in on itself to form a black hole, without giving off any energy.

“The vast majority of simulations show that the easiest way to make this triple work is through direct collapse,” Burdge says.

In addition to giving clues to the black hole’s origins, the outer star has also revealed the system’s age. The physicists observed that the outer star happens to be in the process of becoming a red giant — a phase that occurs at the end of a star’s life. Based on this stellar transition, the team determined that the outer star is about 4 billion years old. Given that neighboring stars are born around the same time, the team concludes that the black hole triple is also 4 billion years old.

“We’ve never been able to do this before for an old black hole,” Burdge says. “Now we know V404 Cygni is part of a triple, it could have formed from direct collapse, and it formed about 4 billion years ago, thanks to this discovery.”

This work was supported, in part, by the National Science Foundation.

Within the human brain, movement is influenced by a brain region called the striatum, which sends instructions to motor neurons in the brain. Those instructions are conveyed by two pathways, one that initiates movement (“go”) and one that suppresses it (“no-go”).

In a new study, MIT researchers have discovered an additional two pathways that arise in the striatum and appear to modulate the effects of the go and no-go pathways. These newly discovered pathways connect to dopamine-producing neurons in the brain — one stimulates dopamine release and the other inhibits it.

By controlling the amount of dopamine in the brain via clusters of neurons known as striosomes, these pathways appear to modify the instructions given by the go and no-go pathways. They may be especially involved in influencing decisions that have a strong emotional component, the researchers say.

“Among all the regions of the striatum, the striosomes alone turned out to be able to project to the dopamine-containing neurons, which we think has something to do with motivation, mood, and controlling movement,” says Ann Graybiel, an MIT Institute Professor, a member of MIT’s McGovern Institute for Brain Research, and the senior author of the new study.

Iakovos Lazaridis, a research scientist at the McGovern Institute, is the lead author of the paper, which appears today in the journal Current Biology.

New pathways

Graybiel has spent much of her career studying the striatum, a structure located deep within the brain that is involved in learning and decision-making, as well as control of movement.

Within the striatum, neurons are arranged in a labyrinth-like structure that includes striosomes, which Graybiel discovered in the 1970s. The classical go and no-go pathways arise from neurons that surround the striosomes, which are known collectively as the matrix. The matrix cells that give rise to these pathways receive input from sensory processing regions such as the visual cortex and auditory cortex. Then, they send go or no-go commands to neurons in the motor cortex.

However, the function of the striosomes, which are not part of those pathways, remained unknown. For many years, researchers in Graybiel’s lab have been trying to solve that mystery.

Their previous work revealed that striosomes receive much of their input from parts of the brain that process emotion. Within striosomes, there are two major types of neurons, classified as D1 and D2. In a 2015 study, Graybiel found that one of these cell types, D1, sends input to the substantia nigra, which is the brain’s major dopamine-producing center.

It took much longer to trace the output of the other set, D2 neurons. In the new Current Biology study, the researchers discovered that those neurons also eventually project to the substantia nigra, but first they connect to a set of neurons in the globus palladus, which inhibits dopamine output. This pathway, an indirect connection to the substantia nigra, reduces the brain’s dopamine output and inhibits movement.

The researchers also confirmed their earlier finding that the pathway arising from D1 striosomes connects directly to the substantia nigra, stimulating dopamine release and initiating movement.

“In the striosomes, we’ve found what is probably a mimic of the classical go/no-go pathways,” Graybiel says. “They’re like classic motor go/no-go pathways, but they don’t go to the motor output neurons of the basal ganglia. Instead, they go to the dopamine cells, which are so important to movement and motivation.”

Emotional decisions

The findings suggest that the classical model of how the striatum controls movement needs to be modified to include the role of these newly identified pathways. The researchers now hope to test their hypothesis that input related to motivation and emotion, which enters the striosomes from the cortex and the limbic system, influences dopamine levels in a way that can encourage or discourage action.

That dopamine release may be especially relevant for actions that induce anxiety or stress. In their 2015 study, Graybiel’s lab found that striosomes play a key role in making decisions that provoke high levels of anxiety; in particular, those that are high risk but may also have a big payoff.

“Ann Graybiel and colleagues have earlier found that the striosome is concerned with inhibiting dopamine neurons. Now they show unexpectedly that another type of striosomal neuron exerts the opposite effect and can signal reward. The striosomes can thus both up- or down-regulate dopamine activity, a very important discovery. Clearly, the regulation of dopamine activity is critical in our everyday life with regard to both movements and mood, to which the striosomes contribute,” says Sten Grillner, a professor of neuroscience at the Karolinska Institute in Sweden, who was not involved in the research.

Another possibility the researchers plan to explore is whether striosomes and matrix cells are arranged in modules that affect motor control of specific parts of the body.

“The next step is trying to isolate some of these modules, and by simultaneously working with cells that belong to the same module, whether they are in the matrix or striosomes, try to pinpoint how the striosomes modulate the underlying function of each of these modules,” Lazaridis says.

They also hope to explore how the striosomal circuits, which project to the same region of the brain that is ravaged by Parkinson’s disease, may influence that disorder.

The research was funded by the National Institutes of Health, the Saks-Kavanaugh Foundation, the William N. and Bernice E. Bumpus Foundation, Jim and Joan Schattinger, the Hock E. Tan and K. Lisa Yang Center for Autism Research, Robert Buxton, the Simons Foundation, the CHDI Foundation, and an Ellen Schapiro and Gerald Axelbaum Investigator BBRF Young Investigator Grant.

Within the human brain, movement is coordinated by a brain region called the striatum, which sends instructions to motor neurons in the brain. Those instructions are conveyed by two pathways, one that initiates movement (“go”) and one that suppresses it (“no-go”).

In a new study, MIT researchers have discovered an additional two pathways that arise in the striatum and appear to modulate the effects of the go and no-go pathways. These newly discovered pathways connect to dopamine-producing neurons in the brain — one stimulates dopamine release and the other inhibits it.

By controlling the amount of dopamine in the brain via clusters of neurons known as striosomes, these pathways appear to modify the instructions given by the go and no-go pathways. They may be especially involved in influencing decisions that have a strong emotional component, the researchers say.

“Among all the regions of the striatum, the striosomes alone turned out to be able to project to the dopamine-containing neurons, which we think has something to do with motivation, mood, and controlling movement,” says Ann Graybiel, an MIT Institute Professor, a member of MIT’s McGovern Institute for Brain Research, and the senior author of the new study.

Iakovos Lazaridis, a research scientist at the McGovern Institute, is the lead author of the paper, which appears today in the journal Current Biology.

New pathways

Graybiel has spent much of her career studying the striatum, a structure located deep within the brain that is involved in learning and decision-making, as well as control of movement.

Within the striatum, neurons are arranged in a labyrinth-like structure that includes striosomes, which Graybiel discovered in the 1970s. The classical go and no-go pathways arise from neurons that surround the striosomes, which are known collectively as the matrix. The matrix cells that give rise to these pathways receive input from sensory processing regions such as the visual cortex and auditory cortex. Then, they send go or no-go commands to neurons in the motor cortex.

However, the function of the striosomes, which are not part of those pathways, remained unknown. For many years, researchers in Graybiel’s lab have been trying to solve that mystery.

Their previous work revealed that striosomes receive much of their input from parts of the brain that process emotion. Within striosomes, there are two major types of neurons, classified as D1 and D2. In a 2015 study, Graybiel found that one of these cell types, D1, sends input to the substantia nigra, which is the brain’s major dopamine-producing center.

It took much longer to trace the output of the other set, D2 neurons. In the new Current Biology study, the researchers discovered that those neurons also eventually project to the substantia nigra, but first they connect to a set of neurons in the globus palladus, which inhibits dopamine output. This pathway, an indirect connection to the substantia nigra, reduces the brain’s dopamine output and inhibits movement.

The researchers also confirmed their earlier finding that the pathway arising from D1 striosomes connects directly to the substantia nigra, stimulating dopamine release and initiating movement.

“In the striosomes, we’ve found what is probably a mimic of the classical go/no-go pathways,” Graybiel says. “They’re like classic motor go/no-go pathways, but they don’t go to the motor output neurons of the basal ganglia. Instead, they go to the dopamine cells, which are so important to movement and motivation.”

Emotional decisions

The findings suggest that the classical model of how the striatum controls movement needs to be modified to include the role of these newly identified pathways. The researchers now hope to test their hypothesis that input related to motivation and emotion, which enters the striosomes from the cortex and the limbic system, influences dopamine levels in a way that can encourage or discourage action.

That dopamine release may be especially relevant for actions that induce anxiety or stress. In their 2015 study, Graybiel’s lab found that striosomes play a key role in making decisions that provoke high levels of anxiety; in particular, those that are high risk but may also have a big payoff.

“Ann Graybiel and colleagues have earlier found that the striosome is concerned with inhibiting dopamine neurons. Now they show unexpectedly that another type of striosomal neuron exerts the opposite effect and can signal reward. The striosomes can thus both up- or down-regulate dopamine activity, a very important discovery. Clearly, the regulation of dopamine activity is critical in our everyday life with regard to both movements and mood, to which the striosomes contribute,” says Sten Grillner, a professor of neuroscience at the Karolinska Institute in Sweden, who was not involved in the research.

Another possibility the researchers plan to explore is whether striosomes and matrix cells are arranged in modules that affect motor control of specific parts of the body.

“The next step is trying to isolate some of these modules, and by simultaneously working with cells that belong to the same module, whether they are in the matrix or striosomes, try to pinpoint how the striosomes modulate the underlying function of each of these modules,” Lazaridis says.

They also hope to explore how the striosomal circuits, which project to the same region of the brain that is ravaged by Parkinson’s disease, may influence that disorder.

The research was funded by the National Institutes of Health, the Saks-Kavanaugh Foundation, the William N. and Bernice E. Bumpus Foundation, Jim and Joan Schattinger, the Hock E. Tan and K. Lisa Yang Center for Autism Research, Robert Buxton, the Simons Foundation, the CHDI Foundation, and an Ellen Schapiro and Gerald Axelbaum Investigator BBRF Young Investigator Grant.

Images of coastal houses being carried off into the sea due to eroding coastlines and powerful storm surges are becoming more commonplace as climate change brings a rising sea level coupled with more powerful storms. In the U.S. alone, coastal storms caused $165 billion in losses in 2022.

Now, a study from MIT shows that protecting and enhancing salt marshes in front of protective seawalls can significantly help protect some coastlines, at a cost that makes this approach reasonable to implement.

The new findings are being reported in the journal Communications Earth and Environment, in a paper by MIT graduate student Ernie I. H. Lee and professor of civil and environmental engineering Heidi Nepf. This study, Nepf says, shows that restoring coastal marshes “is not just something that would be nice to do, but it’s actually economically justifiable.” The researchers found that, among other things, the wave-attenuating effects of salt marsh mean that the seawall behind it can be built significantly lower, reducing construction cost while still providing as much protection from storms.

“One of the other exciting things that the study really brings to light,” Nepf says, “is that you don’t need a huge marsh to get a good effect. It could be a relatively short marsh, just tens of meters wide, that can give you benefit.” That makes her hopeful, Nepf says, that this information might be applied in places where planners may have thought saving a smaller marsh was not worth the expense. “We show that it can make enough of a difference to be financially viable,” she says.

While other studies have previously shown the benefits of natural marshes in attenuating damaging storms, Lee says that such studies “mainly focus on landscapes that have a wide marsh on the order of hundreds of meters. But we want to show that it also applies in urban settings where not as much marsh land is available, especially since in these places existing gray infrastructure (seawalls) tends to already be in place.”

The study was based on computer modeling of waves propagating over different shore profiles, using the morphology of various salt marsh plants — the height and stiffness of the plants, and their spatial density — rather than an empirical drag coefficient. “It’s a physically based model of plant-wave interaction, which allowed us to look at the influence of plant species and changes in morphology across seasons,” without having to go out and calibrate the vegetation drag coefficient with field measurements for each different condition, Nepf says.

The researchers based their benefit-cost analysis on a simple metric: To protect a certain length of shoreline, how much could the height of a given seawall be reduced if it were accompanied by a given amount of marsh? Other ways of assessing the value, such as including the value of real estate that might be damaged by a given amount of flooding, “vary a lot depending on how you value the assets if a flood happens,” Lee says. “We use a more concrete value to quantify the benefits of salt marshes, which is the equivalent height of seawall you would need to deliver the same protection value.”

They used models of a variety of plants, reflecting differences in height and the stiffness across different seasons. They found a twofold variation in the various plants’ effectiveness in attenuating waves, but all provided a useful benefit.

To demonstrate the details in a real-world example and help to validate the simulations, Nepf and Lee studied local salt marshes in Salem, Massachusetts, where projects are already underway to try to restore marshes that had been degraded. Including the specific example provided a template for others, Nepf says. In Salem, their model showed that a healthy salt marsh could offset the need for an additional seawall height of 1.7 meters (about 5.5 feet), based on satisfying a rate of wave overtopping that was set for the safety of pedestrians.

However, the real-world data needed to model a marsh, including maps of salt marsh species, plant height, and shoots per bed area, are “very labor-intensive” to put together, Nepf says. Lee is now developing a method to use drone imaging and machine learning to facilitate this mapmaking. Nepf says this will enable researchers or planners to evaluate a given area of marshland and say, “How much is this marsh worth in terms of its ability to reduce flooding?”

The White House Office of Information and Regulatory Affairs recently released guidance for assessing the value of ecosystem services in planning of federal projects, Nepf explains.  “But in many scenarios, it lacks specific methods for quantifying value, and this study is meeting that need,” she says.

The Federal Emergency Management Agency also has a benefit-cost analysis (BCA) toolkit, Lee notes. “They have guidelines on how to quantify each of the environmental services, and one of the novelties of this paper is quantifying the cost and the protection value of marshes. This is one of the applications that policymakers can consider on how to quantify the environmental service values of marshes,” he says.

The software that environmental engineers can apply to specific sites has been made available online for free on GitHub. “It’s a one-dimensional model accessible by a standard consulting firm,” Nepf says.

“This paper presents a practical tool for translating the wave attenuation capabilities of marshes into economic values, which could assist decision-makers in the adaptation of marshes for nature-based coastal defense,” says Xiaoxia Zhang, an assistant professor at Shenzhen University in China who was not involved in this work. “The results indicate that salt marshes are not only environmentally beneficial but also cost-effective.”

The study “is a very important and crucial step to quantifying the protective value of marshes,” adds Bas Borsje, an associate professor of nature-based flood protection at the University of Twente in the Netherlands, who was not associated with this work. “The most important step missing at the moment is how to translate our findings to the decision makers. This is the first time I’m aware of that decision-makers are quantitatively informed on the protection value of salt marshes.”

Lee received support for this work from the Schoettler Scholarship Fund, administered by the MIT Department of Civil and Environmental Engineering.

It can be hard to connect a certain amount of average global warming with one’s everyday experience, so researchers at MIT have devised a different approach to quantifying the direct impact of climate change. Instead of focusing on global averages, they came up with the concept of “outdoor days”: the number days per year in a given location when the temperature is not too hot or cold to enjoy normal outdoor activities, such as going for a walk, playing sports, working in the garden, or dining outdoors.

In a study published earlier this year, the researchers applied this method to compare the impact of global climate change on different countries around the world, showing that much of the global south would suffer major losses in the number of outdoor days, while some northern countries could see a slight increase. Now, they have applied the same approach to comparing the outcomes for different parts of the United States, dividing the country into nine climatic regions, and finding similar results: Some states, especially Florida and other parts of the Southeast, should see a significant drop in outdoor days, while some, especially in the Northwest, should see a slight increase.

The researchers also looked at correlations between economic activity, such as tourism trends, and changing climate conditions, and examined how numbers of outdoor days could result in significant social and economic impacts. Florida’s economy, for example, is highly dependent on tourism and on people moving there for its pleasant climate; a major drop in days when it is comfortable to spend time outdoors could make the state less of a draw.

The new findings were published this month in the journal Geophysical Research Letters, in a paper by researchers Yeon-Woo Choi and Muhammad Khalifa and professor of civil and environmental engineering Elfatih Eltahir.

“This is something very new in our attempt to understand impacts of climate change impact, in addition to the changing extremes,” Choi says. It allows people to see how these global changes may impact them on a very personal level, as opposed to focusing on global temperature changes or on extreme events such as powerful hurricanes or increased wildfires. “To the best of my knowledge, nobody else takes this same approach” in quantifying the local impacts of climate change, he says. “I hope that many others will parallel our approach to better understand how climate may affect our daily lives.”

The study looked at two different climate scenarios — one where maximum efforts are made to curb global emissions of greenhouse gases and one “worst case” scenario where little is done and global warming continues to accelerate. They used these two scenarios with every available global climate model, 32 in all, and the results were broadly consistent across all 32 models.

The reality may lie somewhere in between the two extremes that were modeled, Eltahir suggests. “I don’t think we’re going to act as aggressively” as the low-emissions scenarios suggest, he says, “and we may not be as careless” as the high-emissions scenario. “Maybe the reality will emerge in the middle, toward the end of the century,” he says.

The team looked at the difference in temperatures and other conditions over various ranges of decades. The data already showed some slight differences in outdoor days from the 1961-1990 period compared to 1991-2020. The researchers then compared these most recent 30 years with the last 30 years of this century, as projected by the models, and found much greater differences ahead for some regions. The strongest effects in the modeling were seen in the Southeastern states. “It seems like climate change is going to have a significant impact on the Southeast in terms of reducing the number of outdoor days,” Eltahir says, “with implications for the quality of life of the population, and also for the attractiveness of tourism and for people who want to retire there.”

He adds that “surprisingly, one of the regions that would benefit a little bit is the Northwest.” But the gain there is modest: an increase of about 14 percent in outdoor days projected for the last three decades of this century, compared to the period from 1976 to 2005. The Southwestern U.S., by comparison, faces an average loss of 23 percent of their outdoor days.

The study also digs into the relationship between climate and economic activity by looking at tourism trends from U.S. National Park Service visitation data, and how that aligned with differences in climate conditions. “Accounting for seasonal variations, we find a clear connection between the number of outdoor days and the number of tourist visits in the United States,” Choi says.

For much of the country, there will be little overall change in the total number of annual outdoor days, the study found, but the seasonal pattern of those days could change significantly. While most parts of the country now see the most outdoor days in summertime, that will shift as summers get hotter, and spring and fall will become the preferred seasons for outdoor activity.

In a way, Eltahir says, “what we are talking about that will happen in the future [for most of the country] is already happening in Florida.” There, he says, “the really enjoyable time of year is in the spring and fall, and summer is not the best time of year.”

People’s level of comfort with temperatures varies somewhat among individuals and among regions, so the researchers designed a tool, now freely available online, that allows people to set their own definitions of the lowest and highest temperatures they consider suitable for outdoor activities, and then see what the climate models predict would be the change in the number of outdoor days for their location, using their own standards of comfort. For their study, they used a widely accepted range of 10 degrees Celsius (50 degrees Fahrenheit) to 25 C (77 F), which is the “thermoneutral zone” in which the human body does not require either metabolic heat generation or evaporative cooling to maintain its core temperature — in other words, in that range there is generally no need to either shiver or sweat.

The model mainly focuses on temperature but also allows people to include humidity or precipitation in their definition of what constitutes a comfortable outdoor day. The model could be extended to incorporate other variables such as air quality, but the researchers say temperature tends to be the major determinant of comfort for most people.

Using their software tool, “If you disagree with how we define an outdoor day, you could define one for yourself, and then you’ll see what the impacts of that are on your number of outdoor days and their seasonality,” Eltahir says.

This work was inspired by the realization, he says, that “people’s understanding of climate change is based on the assumption that climate change is something that’s going to happen sometime in the future and going to happen to someone else. It’s not going to impact them directly. And I think that contributes to the fact that we are not doing enough.”

Instead, the concept of outdoor days “brings the concept of climate change home, brings it to personal everyday activities,” he says. “I hope that people will find that useful to bridge that gap, and provide a better understanding and appreciation of the problem. And hopefully that would help lead to sound policies that are based on science, regarding climate change.”

The research was based on work supported by the Community Jameel for Jameel Observatory CREWSnet and Abdul Latif Jameel Water and Food Systems Lab at MIT.

Despite their impressive capabilities, large language models are far from perfect. These artificial intelligence models sometimes “hallucinate” by generating incorrect or unsupported information in response to a query.

Due to this hallucination problem, an LLM’s responses are often verified by human fact-checkers, especially if a model is deployed in a high-stakes setting like health care or finance. However, validation processes typically require people to read through long documents cited by the model, a task so onerous and error-prone it may prevent some users from deploying generative AI models in the first place.

To help human validators, MIT researchers created a user-friendly system that enables people to verify an LLM’s responses much more quickly. With this tool, called SymGen, an LLM generates responses with citations that point directly to the place in a source document, such as a given cell in a database.

Users hover over highlighted portions of its text response to see data the model used to generate that specific word or phrase. At the same time, the unhighlighted portions show users which phrases need additional attention to check and verify.

“We give people the ability to selectively focus on parts of the text they need to be more worried about. In the end, SymGen can give people higher confidence in a model’s responses because they can easily take a closer look to ensure that the information is verified,” says Shannon Shen, an electrical engineering and computer science graduate student and co-lead author of a paper on SymGen.

Through a user study, Shen and his collaborators found that SymGen sped up verification time by about 20 percent, compared to manual procedures. By making it faster and easier for humans to validate model outputs, SymGen could help people identify errors in LLMs deployed in a variety of real-world situations, from generating clinical notes to summarizing financial market reports.

Shen is joined on the paper by co-lead author and fellow EECS graduate student Lucas Torroba Hennigen; EECS graduate student Aniruddha “Ani” Nrusimha; Bernhard Gapp, president of the Good Data Initiative; and senior authors David Sontag, a professor of EECS, a member of the MIT Jameel Clinic, and the leader of the Clinical Machine Learning Group of the Computer Science and Artificial Intelligence Laboratory (CSAIL); and Yoon Kim, an assistant professor of EECS and a member of CSAIL. The research was recently presented at the Conference on Language Modeling.

Symbolic references

To aid in validation, many LLMs are designed to generate citations, which point to external documents, along with their language-based responses so users can check them. However, these verification systems are usually designed as an afterthought, without considering the effort it takes for people to sift through numerous citations, Shen says.

“Generative AI is intended to reduce the user’s time to complete a task. If you need to spend hours reading through all these documents to verify the model is saying something reasonable, then it’s less helpful to have the generations in practice,” Shen says.

The researchers approached the validation problem from the perspective of the humans who will do the work.

A SymGen user first provides the LLM with data it can reference in its response, such as a table that contains statistics from a basketball game. Then, rather than immediately asking the model to complete a task, like generating a game summary from those data, the researchers perform an intermediate step. They prompt the model to generate its response in a symbolic form.

With this prompt, every time the model wants to cite words in its response, it must write the specific cell from the data table that contains the information it is referencing. For instance, if the model wants to cite the phrase “Portland Trailblazers” in its response, it would replace that text with the cell name in the data table that contains those words.

“Because we have this intermediate step that has the text in a symbolic format, we are able to have really fine-grained references. We can say, for every single span of text in the output, this is exactly where in the data it corresponds to,” Torroba Hennigen says.

SymGen then resolves each reference using a rule-based tool that copies the corresponding text from the data table into the model’s response.

“This way, we know it is a verbatim copy, so we know there will not be any errors in the part of the text that corresponds to the actual data variable,” Shen adds.

Streamlining validation

The model can create symbolic responses because of how it is trained. Large language models are fed reams of data from the internet, and some data are recorded in “placeholder format” where codes replace actual values.

When SymGen prompts the model to generate a symbolic response, it uses a similar structure.

“We design the prompt in a specific way to draw on the LLM’s capabilities,” Shen adds.

During a user study, the majority of participants said SymGen made it easier to verify LLM-generated text. They could validate the model’s responses about 20 percent faster than if they used standard methods.

However, SymGen is limited by the quality of the source data. The LLM could cite an incorrect variable, and a human verifier may be none-the-wiser.

In addition, the user must have source data in a structured format, like a table, to feed into SymGen. Right now, the system only works with tabular data.

Moving forward, the researchers are enhancing SymGen so it can handle arbitrary text and other forms of data. With that capability, it could help validate portions of AI-generated legal document summaries, for instance. They also plan to test SymGen with physicians to study how it could identify errors in AI-generated clinical summaries.

This work is funded, in part, by Liberty Mutual and the MIT Quest for Intelligence Initiative.

The much-touted arrival of “precision medicine” promises tailored technologies that help individuals and may also reduce health care costs. New research shows how pregnancy screening can meet both of these objectives, but the findings also highlight how precision medicine must be matched well with patients to save money.

The study involves cfDNA screenings, a type of blood test that can reveal conditions based on chromosomal variation, such as Down Syndrome. For many pregnant women, though not all, cfDNA screenings can be an alternative to amniocentesis or chorionic villus sampling (CVS) — invasive procedures that come with a risk of miscarriage.

In examining how widely cfDNA tests should be used, the study reached a striking conclusion.

“What we find is the highest value for the cfDNA testing comes from people who are high risk, but not extraordinarily high risk,” says Amy Finkelstein, an MIT economist and co-author of a newly published paper detailing the study.

The paper, “Targeting Precision Medicine: Evidence from Prenatal Screening,” appears in the Journal of Political Economy. The co-authors are Peter Conner, an associate professor and senior consultant at Karolinska University Hospital in Sweden; Liran Einav, a professor of economics at Stanford University; Finkelstein, the John and Jennie S. MacDonald Professor of Economics at MIT; and Petra Persson, an assistant professor of economics at Stanford University.

“There is a lot of hope attached to precision medicine,” Persson says. “We can do a lot of new things and tailor health care treatments to patients, which holds a lot of promise. In this paper, we highlight that while this is all true, there are also significant costs in the personalization of medicine. As a society, we may want to examine how to use these technologies while keeping an eye on health care costs.”

Measuring the benefit to “middle-risk” patients

To conduct the study, the research team looked at the introduction of cfDNA screening in Sweden, during the period from 2011 to 2019, with data covering over 230,000 pregnancies. As it happens, there were also regional discrepancies in the extent to which cfDNA screenings were covered by Swedish health care, for patients not already committed to having invasive testing. Some regions covered cfDNA testing quite widely, for all patients with a “moderate” assessed risk or higher; other regions, by contrast, restricted coverage to a subset of patients within that group with elevated risk profiles. This provided variation the researchers could use when conducting their analysis.

With the most generous coverage of cfDNA testing, the procedure was used by 86 percent of patients; with more targeted coverage, that figure dropped to about 33 percent. In both cases, the amount of invasive testing, including amniocentesis, dropped significantly, to about 5 percent. (The cfDNA screenings are very informative, but not fully conclusive, which invasive testing is, so some pregnant women will opt-for a follow-up procedure.)

Both approaches, then, yielded similar reductions in the rate of invasive testing. But due to the costs of cfDNA tests, the economic implications are quite different. Introducing wide coverage of cfDNA tests would raise overall medical costs by about $250 per pregnancy, the study estimates. In contrast, introducing cfDNA with more targeted coverage yields a reduction of about $89 per patient.

Ultimately, the larger dynamics are clear. Pregnant women who have the highest risk of bearing children with chromosome-based conditions are likely to still opt for an invasive test like amniocentesis. Those with virtually no risk may not even have cfDNA tests done. For a group in between, cfDNA tests have a substantial medical value, relieving them of the need for an invasive test. And narrowing the group of patients getting cfDNA tests lowers the overall cost.

“People who are very high-risk are often going to use the invasive test, which is definitive, regardless of whether they have a cfDNA screen or not,” Finkelstein says. “But for middle-risk people, covering cfDNA produces a big increase in cfDNA testing, and that produces a big decline in the rates of the riskier, and more expensive, invasive test.”

How precise?

In turn, the study’s findings raise a larger point. Precision medicine, in almost any form, will add expenses to medical care. Therefore developing some precision about who receives it is significant.

“The allure of precision medicine is targeting people who need it, so we don’t do expensive and potentially unpleasant tests and treatments of people who don’t need them,” Finkelstein says. “Which sounds great, but it kicks the can down the road. You still need to figure out who is a candidate for which kind of precision medicine.”

Therefore, in medicine, instead of just throwing technology at the problem, we may want to aim carefully, where evidence warrants it. Overall, that means good precision medicine builds on good policy analysis, not just good technology.

“Sometimes when we think medical technology has an impact, we simply ask if the technology raises or lowers health care costs, or if it makes patients healthier,” Persson observes. “An important insight from our work, I think, is that the answers are not just about the technology. It’s about the pairing of technology and policy because policy is going to influence the impact of technology on health care and patient outcomes. We see this clearly in our study.”

In this case, finding comparable patient outcomes with narrower cfDNA screenings suggests one way of targeting diagnostic procedures. And across many possible medical situations, finding the subset of people for whom a technology is most likely to yield new and actionable information seems a promising objective.

“The benefit is not just an innate feature of the testing,” Finkelstein says. “With diagnostic technologies, the value of information is greatest when you’re neither obviously appropriate or inappropriate for the next treatment. It’s really the non-monotone value of information that’s interesting.”

The study was supported, in part, by the U.S. National Science Foundation.

Some of the most widely used drugs today, including penicillin, were discovered through a process called phenotypic screening. Using this method, scientists are essentially throwing drugs at a problem — for example, when attempting to stop bacterial growth or fixing a cellular defect — and then observing what happens next, without necessarily first knowing how the drug works. Perhaps surprisingly, historical data show that this approach is better at yielding approved medicines than those investigations that more narrowly focus on specific molecular targets.

But many scientists believe that properly setting up the problem is the true key to success. Certain microbial infections or genetic disorders caused by single mutations are much simpler to prototype than complex diseases like cancer. These require intricate biological models that are far harder to make or acquire. The result is a bottleneck in the number of drugs that can be tested, and thus the usefulness of phenotypic screening.

Now, a team of scientists led by the Shalek Lab at MIT has developed a promising new way to address the difficulty of applying phenotyping screening to scale. Their method allows researchers to simultaneously apply multiple drugs to a biological problem at once, and then computationally work backward to figure out the individual effects of each. For instance, when the team applied this method to models of pancreatic cancer and human immune cells, they were able to uncover surprising new biological insights, while also minimizing cost and sample requirements by several-fold — solving a few problems in scientific research at once.

Zev Gartner, a professor in pharmaceutical chemistry at the University of California at San Francisco, says this new method has great potential. “I think if there is a strong phenotype one is interested in, this will be a very powerful approach,” Gartner says.

The research was published Oct. 8 in Nature Biotechnology. It was led by Ivy Liu, Walaa Kattan, Benjamin Mead, Conner Kummerlowe, and Alex K. Shalek, the director of the Institute for Medical Engineering and Sciences (IMES) and the Health Innovation Hub at MIT, as well as the J. W. Kieckhefer Professor in IMES and the Department of Chemistry. It was supported by the National Institutes of Health and the Bill and Melinda Gates Foundation.

A “crazy” way to increase scale

Technological advances over the past decade have revolutionized our understanding of the inner lives of individual cells, setting the stage for richer phenotypic screens. However, many challenges remain.

For one, biologically representative models like organoids and primary tissues are only available in limited quantities. The most informative tests, like single-cell RNA sequencing, are also expensive, time-consuming, and labor-intensive.

That’s why the team decided to test out the “bold, maybe even crazy idea” to mix everything together, says Liu, a PhD student in the MIT Computational and Systems Biology program. In other words, they chose to combine many perturbations — things like drugs, chemical molecules, or biological compounds made by cells — into one single concoction, and then try to decipher their individual effects afterward.

They began testing their workflow by making different combinations of 316 U.S. Food and Drug Administration-approved drugs. “It’s a high bar: basically, the worst-case scenario,” says Liu. “Since every drug is known to have a strong effect, the signals could have been impossible to disentangle.”

These random combinations ranged from three to 80 drugs per pool, each of which was applied to lab-grown cells. The team then tried to understand the effects of the individual drug using a linear computational model.

It was a success. When compared with traditional tests for each individual drug, the new method yielded comparable results, successfully finding the strongest drugs and their respective effects in each pool, at a fraction of the cost, samples, and effort.

Putting it into practice

To test the method’s applicability to address real-world health challenges, the team then approached two problems that were previously unimaginable with past phenotypic screening techniques.

The first test focused on pancreatic ductal adenocarcinoma (PDAC), one of the deadliest types of cancer. In PDAC, many types of signals come from the surrounding cells in the tumor's environment. These signals can influence how the tumor progresses and responds to treatments. So, the team wanted to identify the most important ones.

Using their new method to pool different signals in parallel, they found several surprise candidates. “We never could have predicted some of our hits,” says Shalek. These included two previously overlooked cytokines that actually could predict survival outcomes of patients with PDAC in public cancer data sets.

The second test looked at the effects of 90 drugs on adjusting the immune system’s function. These drugs were applied to fresh human blood cells, which contain a complex mix of different types of immune cells. Using their new method and single-cell RNA-sequencing, the team could not only test a large library of drugs, but also separate the drugs’ effects out for each type of cell. This enabled the team to understand how each drug might work in a more complex tissue, and then select the best one for the job.

“We might say there’s a defect in a T cell, so we’re going to add this drug, but we never think about, well, what does that drug do to all of the other cells in the tissue?” says Shalek. “We now have a way to gather this information, so that we can begin to pick drugs to maximize on-target effects and minimize side effects.”

Together, these experiments also showed Shalek the need to build better tools and datasets for creating hypotheses about potential treatments. “The complexity and lack of predictability for the responses we saw tells me that we likely are not finding the right, or most effective, drugs in many instances,” says Shalek.

Reducing barriers and improving lives

Although the current compression technique can identify the perturbations with the greatest effects, it’s still unable to perfectly resolve the effects of each one. Therefore, the team recommends that it act as a supplement to support additional screening. “Traditional tests that examine the top hits should follow,” Liu says.

Importantly, however, the new compression framework drastically reduces the number of input samples, costs, and labor required to execute a screen. With fewer barriers in play, it marks an exciting advance for understanding complex responses in different cells and building new models for precision medicine.

Shalek says, “This is really an incredible approach that opens up the kinds of things that we can do to find the right targets, or the right drugs, to use to improve lives for patients.”

A quasar is the extremely bright core of a galaxy that hosts an active supermassive black hole at its center. As the black hole draws in surrounding gas and dust, it blasts out an enormous amount of energy, making quasars some of the brightest objects in the universe. Quasars have been observed as early as a few hundred million years after the Big Bang, and it’s been a mystery as to how these objects could have grown so bright and massive in such a short amount of cosmic time.

Scientists have proposed that the earliest quasars sprang from overly dense regions of primordial matter, which would also have produced many smaller galaxies in the quasars’ environment. But in a new MIT-led study, astronomers observed some ancient quasars that appear to be surprisingly alone in the early universe.

The astronomers used NASA’s James Webb Space Telescope (JWST) to peer back in time, more than 13 billion years, to study the cosmic surroundings of five known ancient quasars. They found a surprising variety in their neighborhoods, or “quasar fields.” While some quasars reside in very crowded fields with more than 50 neighboring galaxies, as all models predict, the remaining quasars appear to drift in voids, with only a few stray galaxies in their vicinity.

These lonely quasars are challenging physicists’ understanding of how such luminous objects could have formed so early on in the universe, without a significant source of surrounding matter to fuel their black hole growth.

“Contrary to previous belief, we find on average, these quasars are not necessarily in those highest-density regions of the early universe. Some of them seem to be sitting in the middle of nowhere,” says Anna-Christina Eilers, assistant professor of physics at MIT. “It’s difficult to explain how these quasars could have grown so big if they appear to have nothing to feed from.”

There is a possibility that these quasars may not be as solitary as they appear, but are instead surrounded by galaxies that are heavily shrouded in dust and therefore hidden from view. Eilers and her colleagues hope to tune their observations to try and see through any such cosmic dust, in order to understand how quasars grew so big, so fast, in the early universe.

Eilers and her colleagues report their findings in a paper appearing today in the Astrophysical Journal. The MIT co-authors include postdocs Rohan Naidu and Minghao Yue; Robert Simcoe, the Francis Friedman Professor of Physics and director of MIT’s Kavli Institute for Astrophysics and Space Research; and collaborators from institutions including Leiden University, the University of California at Santa Barbara, ETH Zurich, and elsewhere.

Galactic neighbors

The five newly observed quasars are among the oldest quasars observed to date. More than 13 billion years old, the objects are thought to have formed between 600 to 700 million years after the Big Bang. The supermassive black holes powering the quasars are a billion times more massive than the sun, and more than a trillion times brighter. Due to their extreme luminosity, the light from each quasar is able to travel over the age of the universe, far enough to reach JWST’s highly sensitive detectors today.

“It’s just phenomenal that we now have a telescope that can capture light from 13 billion years ago in so much detail,” Eilers says. “For the first time, JWST enabled us to look at the environment of these quasars, where they grew up, and what their neighborhood was like.”

The team analyzed images of the five ancient quasars taken by JWST between August 2022 and June 2023. The observations of each quasar comprised multiple “mosaic” images, or partial views of the quasar’s field, which the team effectively stitched together to produce a complete picture of each quasar’s surrounding neighborhood.

The telescope also took measurements of light in multiple wavelengths across each quasar’s field, which the team then processed to determine whether a given object in the field was light from a neighboring galaxy, and how far a galaxy is from the much more luminous central quasar.

“We found that the only difference between these five quasars is that their environments look so different,” Eilers says. “For instance, one quasar has almost 50 galaxies around it, while another has just two. And both quasars are within the same size, volume, brightness, and time of the universe. That was really surprising to see.”

Growth spurts

The disparity in quasar fields introduces a kink in the standard picture of black hole growth and galaxy formation. According to physicists’ best understanding of how the first objects in the universe emerged, a cosmic web of dark matter should have set the course. Dark matter is an as-yet unknown form of matter that has no other interactions with its surroundings other than through gravity.

Shortly after the Big Bang, the early universe is thought to have formed filaments of dark matter that acted as a sort of gravitational road, attracting gas and dust along its tendrils. In overly dense regions of this web, matter would have accumulated to form more massive objects. And the brightest, most massive early objects, such as quasars, would have formed in the web’s highest-density regions, which would have also churned out many more, smaller galaxies.

“The cosmic web of dark matter is a solid prediction of our cosmological model of the Universe, and it can be described in detail using numerical simulations,” says co-author Elia Pizzati, a graduate student at Leiden University. “By comparing our observations to these simulations, we can determine where in the cosmic web quasars are located.”

Scientists estimate that quasars would have had to grow continuously with very high accretion rates in order to reach the extreme mass and luminosities at the times that astronomers have observed them, fewer than 1 billion years after the Big Bang.

“The main question we’re trying to answer is, how do these billion-solar-mass black holes form at a time when the universe is still really, really young? It’s still in its infancy,” Eilers says.

The team’s findings may raise more questions than answers. The “lonely” quasars appear to live in relatively empty regions of space. If physicists’ cosmological models are correct, these barren regions signify very little dark matter, or starting material for brewing up stars and galaxies. How, then, did extremely bright and massive quasars come to be?

“Our results show that there’s still a significant piece of the puzzle missing of how these supermassive black holes grow,” Eilers says. “If there’s not enough material around for some quasars to be able to grow continuously, that means there must be some other way that they can grow, that we have yet to figure out.”

This research was supported, in part, by the European Research Council. 

In the current AI zeitgeist, sequence models have skyrocketed in popularity for their ability to analyze data and predict what to do next. For instance, you’ve likely used next-token prediction models like ChatGPT, which anticipate each word (token) in a sequence to form answers to users’ queries. There are also full-sequence diffusion models like Sora, which convert words into dazzling, realistic visuals by successively “denoising” an entire video sequence. 

Researchers from MIT’s Computer Science and Artificial Intelligence Laboratory (CSAIL) have proposed a simple change to the diffusion training scheme that makes this sequence denoising considerably more flexible.

When applied to fields like computer vision and robotics, the next-token and full-sequence diffusion models have capability trade-offs. Next-token models can spit out sequences that vary in length. However, they make these generations while being unaware of desirable states in the far future — such as steering its sequence generation toward a certain goal 10 tokens away — and thus require additional mechanisms for long-horizon (long-term) planning. Diffusion models can perform such future-conditioned sampling, but lack the ability of next-token models to generate variable-length sequences.

Researchers from CSAIL want to combine the strengths of both models, so they created a sequence model training technique called “Diffusion Forcing.” The name comes from “Teacher Forcing,” the conventional training scheme that breaks down full sequence generation into the smaller, easier steps of next-token generation (much like a good teacher simplifying a complex concept).

Diffusion Forcing found common ground between diffusion models and teacher forcing: They both use training schemes that involve predicting masked (noisy) tokens from unmasked ones. In the case of diffusion models, they gradually add noise to data, which can be viewed as fractional masking. The MIT researchers’ Diffusion Forcing method trains neural networks to cleanse a collection of tokens, removing different amounts of noise within each one while simultaneously predicting the next few tokens. The result: a flexible, reliable sequence model that resulted in higher-quality artificial videos and more precise decision-making for robots and AI agents.

By sorting through noisy data and reliably predicting the next steps in a task, Diffusion Forcing can aid a robot in ignoring visual distractions to complete manipulation tasks. It can also generate stable and consistent video sequences and even guide an AI agent through digital mazes. This method could potentially enable household and factory robots to generalize to new tasks and improve AI-generated entertainment.

“Sequence models aim to condition on the known past and predict the unknown future, a type of binary masking. However, masking doesn’t need to be binary,” says lead author, MIT electrical engineering and computer science (EECS) PhD student, and CSAIL member Boyuan Chen. “With Diffusion Forcing, we add different levels of noise to each token, effectively serving as a type of fractional masking. At test time, our system can “unmask” a collection of tokens and diffuse a sequence in the near future at a lower noise level. It knows what to trust within its data to overcome out-of-distribution inputs.”

In several experiments, Diffusion Forcing thrived at ignoring misleading data to execute tasks while anticipating future actions.

When implemented into a robotic arm, for example, it helped swap two toy fruits across three circular mats, a minimal example of a family of long-horizon tasks that require memories. The researchers trained the robot by controlling it from a distance (or teleoperating it) in virtual reality. The robot is trained to mimic the user’s movements from its camera. Despite starting from random positions and seeing distractions like a shopping bag blocking the markers, it placed the objects into its target spots.

To generate videos, they trained Diffusion Forcing on “Minecraft” game play and colorful digital environments created within Google’s DeepMind Lab Simulator. When given a single frame of footage, the method produced more stable, higher-resolution videos than comparable baselines like a Sora-like full-sequence diffusion model and ChatGPT-like next-token models. These approaches created videos that appeared inconsistent, with the latter sometimes failing to generate working video past just 72 frames.

Diffusion Forcing not only generates fancy videos, but can also serve as a motion planner that steers toward desired outcomes or rewards. Thanks to its flexibility, Diffusion Forcing can uniquely generate plans with varying horizon, perform tree search, and incorporate the intuition that the distant future is more uncertain than the near future. In the task of solving a 2D maze, Diffusion Forcing outperformed six baselines by generating faster plans leading to the goal location, indicating that it could be an effective planner for robots in the future.

Across each demo, Diffusion Forcing acted as a full sequence model, a next-token prediction model, or both. According to Chen, this versatile approach could potentially serve as a powerful backbone for a “world model,” an AI system that can simulate the dynamics of the world by training on billions of internet videos. This would allow robots to perform novel tasks by imagining what they need to do based on their surroundings. For example, if you asked a robot to open a door without being trained on how to do it, the model could produce a video that’ll show the machine how to do it.

The team is currently looking to scale up their method to larger datasets and the latest transformer models to improve performance. They intend to broaden their work to build a ChatGPT-like robot brain that helps robots perform tasks in new environments without human demonstration.

“With Diffusion Forcing, we are taking a step to bringing video generation and robotics closer together,” says senior author Vincent Sitzmann, MIT assistant professor and member of CSAIL, where he leads the Scene Representation group. “In the end, we hope that we can use all the knowledge stored in videos on the internet to enable robots to help in everyday life. Many more exciting research challenges remain, like how robots can learn to imitate humans by watching them even when their own bodies are so different from our own!”

Chen and Sitzmann wrote the paper alongside recent MIT visiting researcher Diego Martí Monsó, and CSAIL affiliates: Yilun Du, a EECS graduate student; Max Simchowitz, former postdoc and incoming Carnegie Mellon University assistant professor; and Russ Tedrake, the Toyota Professor of EECS, Aeronautics and Astronautics, and Mechanical Engineering at MIT, vice president of robotics research at the Toyota Research Institute, and CSAIL member. Their work was supported, in part, by the U.S. National Science Foundation, the Singapore Defence Science and Technology Agency, Intelligence Advanced Research Projects Activity via the U.S. Department of the Interior, and the Amazon Science Hub. They will present their research at NeurIPS in December.

When an election result is disputed, people who are skeptical about the outcome may be swayed by figures of authority who come down on one side or the other. Those figures can be independent monitors, political figures, or news organizations. However, these “debunking” efforts don’t always have the desired effect, and in some cases, they can lead people to cling more tightly to their original position.

Neuroscientists and political scientists at MIT and the University of California at Berkeley have now created a computational model that analyzes the factors that help to determine whether debunking efforts will persuade people to change their beliefs about the legitimacy of an election. Their findings suggest that while debunking fails much of the time, it can be successful under the right conditions.

For instance, the model showed that successful debunking is more likely if people are less certain of their original beliefs and if they believe the authority is unbiased or strongly motivated by a desire for accuracy. It also helps when an authority comes out in support of a result that goes against a bias they are perceived to hold: for example, Fox News declaring that Joseph R. Biden had won in Arizona in the 2020 U.S. presidential election.

“When people see an act of debunking, they treat it as a human action and understand it the way they understand human actions — that is, as something somebody did for their own reasons,” says Rebecca Saxe, the John W. Jarve Professor of Brain and Cognitive Sciences, a member of MIT’s McGovern Institute for Brain Research, and the senior author of the study. “We’ve used a very simple, general model of how people understand other people’s actions, and found that that’s all you need to describe this complex phenomenon.”

The findings could have implications as the United States prepares for the presidential election taking place on Nov. 5, as they help to reveal the conditions that would be most likely to result in people accepting the election outcome.

MIT graduate student Setayesh Radkani is the lead author of the paper, which appears today in a special election-themed issue of the journal PNAS Nexus. Marika Landau-Wells PhD ’18, a former MIT postdoc who is now an assistant professor of political science at the University of California at Berkeley, is also an author of the study.

Modeling motivation

In their work on election debunking, the MIT team took a novel approach, building on Saxe’s extensive work studying “theory of mind” — how people think about the thoughts and motivations of other people.

As part of her PhD thesis, Radkani has been developing a computational model of the cognitive processes that occur when people see others being punished by an authority. Not everyone interprets punitive actions the same way, depending on their previous beliefs about the action and the authority. Some may see the authority as acting legitimately to punish an act that was wrong, while others may see an authority overreaching to issue an unjust punishment.

Last year, after participating in an MIT workshop on the topic of polarization in societies, Saxe and Radkani had the idea to apply the model to how people react to an authority attempting to sway their political beliefs. They enlisted Landau-Wells, who received her PhD in political science before working as a postdoc in Saxe’s lab, to join their effort, and Landau suggested applying the model to debunking of beliefs regarding the legitimacy of an election result.

The computational model created by Radkani is based on Bayesian inference, which allows the model to continually update its predictions of people’s beliefs as they receive new information. This approach treats debunking as an action that a person undertakes for his or her own reasons. People who observe the authority’s statement then make their own interpretation of why the person said what they did. Based on that interpretation, people may or may not change their own beliefs about the election result.

Additionally, the model does not assume that any beliefs are necessarily incorrect or that any group of people is acting irrationally.

“The only assumption that we made is that there are two groups in the society that differ in their perspectives about a topic: One of them thinks that the election was stolen and the other group doesn’t,” Radkani says. “Other than that, these groups are similar. They share their beliefs about the authority — what the different motives of the authority are and how motivated the authority is by each of those motives.”

The researchers modeled more than 200 different scenarios in which an authority attempts to debunk a belief held by one group regarding the validity of an election outcome.

Each time they ran the model, the researchers altered the certainty levels of each group’s original beliefs, and they also varied the groups’ perceptions of the motivations of the authority. In some cases, groups believed the authority was motivated by promoting accuracy, and in others they did not. The researchers also altered the groups’ perceptions of whether the authority was biased toward a particular viewpoint, and how strongly the groups believed in those perceptions.

Building consensus

In each scenario, the researchers used the model to predict how each group would respond to a series of five statements made by an authority trying to convince them that the election had been legitimate. The researchers found that in most of the scenarios they looked at, beliefs remained polarized and in some cases became even further polarized. This polarization could also extend to new topics unrelated to the original context of the election, the researchers found.

However, under some circumstances, the debunking was successful, and beliefs converged on an accepted outcome. This was more likely to happen when people were initially more uncertain about their original beliefs.

“When people are very, very certain, they become hard to move. So, in essence, a lot of this authority debunking doesn’t matter,” Landau-Wells says. “However, there are a lot of people who are in this uncertain band. They have doubts, but they don’t have firm beliefs. One of the lessons from this paper is that we’re in a space where the model says you can affect people’s beliefs and move them towards true things.”

Another factor that can lead to belief convergence is if people believe that the authority is unbiased and highly motivated by accuracy. Even more persuasive is when an authority makes a claim that goes against their perceived bias — for instance, Republican governors stating that elections in their states had been fair even though the Democratic candidate won.

As the 2024 presidential election approaches, grassroots efforts have been made to train nonpartisan election observers who can vouch for whether an election was legitimate. These types of organizations may be well-positioned to help sway people who might have doubts about the election’s legitimacy, the researchers say.

“They’re trying to train to people to be independent, unbiased, and committed to the truth of the outcome more than anything else. Those are the types of entities that you want. We want them to succeed in being seen as independent. We want them to succeed as being seen as truthful, because in this space of uncertainty, those are the voices that can move people toward an accurate outcome,” Landau-Wells says.

The research was funded, in part, by the Patrick J. McGovern Foundation and the Guggenheim Foundation.

Active electronics — components that can control electrical signals — usually contain semiconductor devices that receive, store, and process information. These components, which must be made in a clean room, require advanced fabrication technology that is not widely available outside a few specialized manufacturing centers.

During the Covid-19 pandemic, the lack of widespread semiconductor fabrication facilities was one cause of a worldwide electronics shortage, which drove up costs for consumers and had implications in everything from economic growth to national defense. The ability to 3D print an entire, active electronic device without the need for semiconductors could bring electronics fabrication to businesses, labs, and homes across the globe.

While this idea is still far off, MIT researchers have taken an important step in that direction by demonstrating fully 3D-printed resettable fuses, which are key components of active electronics that usually require semiconductors.

The researchers’ semiconductor-free devices, which they produced using standard 3D printing hardware and an inexpensive, biodegradable material, can perform the same switching functions as the semiconductor-based transistors used for processing operations in active electronics.

Although still far from achieving the performance of semiconductor transistors, the 3D-printed devices could be used for basic control operations like regulating the speed of an electric motor.

“This technology has real legs. While we cannot compete with silicon as a semiconductor, our idea is not to necessarily replace what is existing, but to push 3D printing technology into uncharted territory. In a nutshell, this is really about democratizing technology. This could allow anyone to create smart hardware far from traditional manufacturing centers,” says Luis Fernando Velásquez-García, a principal research scientist in MIT’s Microsystems Technology Laboratories (MTL) and senior author of a paper describing the devices, which appears in Virtual and Physical Prototyping.

He is joined on the paper by lead author Jorge Cañada, an electrical engineering and computer science graduate student.

An unexpected project

Semiconductors, including silicon, are materials with electrical properties that can be tailored by adding certain impurities. A silicon device can have conductive and insulating regions, depending on how it is engineered. These properties make silicon ideal for producing transistors, which are a basic building block of modern electronics.

However, the researchers didn’t set out to 3D-print semiconductor-free devices that could behave like silicon-based transistors.

This project grew out of another in which they were fabricating magnetic coils using extrusion printing, a process where the printer melts filament and squirts material through a nozzle, fabricating an object layer-by-layer.

They saw an interesting phenomenon in the material they were using, a polymer filament doped with copper nanoparticles.

If they passed a large amount of electric current into the material, it would exhibit a huge spike in resistance but would return to its original level shortly after the current flow stopped.

This property enables engineers to make transistors that can operate as switches, something that is typically only associated with silicon and other semiconductors. Transistors, which switch on and off to process binary data, are used to form logic gates which perform computation.

“We saw that this was something that could help take 3D printing hardware to the next level. It offers a clear way to provide some degree of ‘smart’ to an electronic device,” Velásquez-García says.

The researchers tried to replicate the same phenomenon with other 3D printing filaments, testing polymers doped with carbon, carbon nanotubes, and graphene. In the end, they could not find another printable material that could function as a resettable fuse.

They hypothesize that the copper particles in the material spread out when it is heated by the electric current, which causes a spike in resistance that comes back down when the material cools and the copper particles move closer together. They also think the polymer base of the material changes from crystalline to amorphous when heated, then returns to crystalline when cooled down — a phenomenon known as the polymeric positive temperature coefficient.

“For now, that is our best explanation, but that is not the full answer because that doesn’t explain why it only happened in this combination of materials. We need to do more research, but there is no doubt that this phenomenon is real,” he says.

3D-printing active electronics

The team leveraged the phenomenon to print switches in a single step that could be used to form semiconductor-free logic gates.

The devices are made from thin, 3D-printed traces of the copper-doped polymer. They contain intersecting conductive regions that enable the researchers to regulate the resistance by controlling the voltage fed into the switch.

While the devices did not perform as well as silicon-based transistors, they could be used for simpler control and processing functions, such as turning a motor on and off. Their experiments showed that, even after 4,000 cycles of switching, the devices showed no signs of deterioration.

But there are limits to how small the researchers can make the switches, based on the physics of extrusion printing and the properties of the material. They could print devices that were a few hundred microns, but transistors in state-of-the-art electronics are only few nanometers in diameter.

“The reality is that there are many engineering situations that don’t require the best chips. At the end of the day, all you care about is whether your device can do the task. This technology is able to satisfy a constraint like that,” he says.

However, unlike semiconductor fabrication, their technique uses a biodegradable material and the process uses less energy and produces less waste. The polymer filament could also be doped with other materials, like magnetic microparticles that could enable additional functionalities.

In the future, the researchers want to use this technology to print fully functional electronics. They are striving to fabricate a working magnetic motor using only extrusion 3D printing. They also want to finetune the process so they could build more complex circuits and see how far they can push the performance of these devices.

“This paper demonstrates that active electronic devices can be made using extruded polymeric conductive materials. This technology enables electronics to be built into 3D printed structures. An intriguing application is on-demand 3D printing of mechatronics on board spacecraft,” says Roger Howe, the William E. Ayer Professor of Engineering, Emeritus, at Stanford University, who was not involved with this work.

This work is funded, in part, by Empiriko Corporation.

A classical way to image nanoscale structures in cells is with high-powered, expensive super-resolution microscopes. As an alternative, MIT researchers have developed a way to expand tissue before imaging it — a technique that allows them to achieve nanoscale resolution with a conventional light microscope.

In the newest version of this technique, the researchers have made it possible to expand tissue 20-fold in a single step. This simple, inexpensive method could pave the way for nearly any biology lab to perform nanoscale imaging.

“This democratizes imaging,” says Laura Kiessling, the Novartis Professor of Chemistry at MIT and a member of the Broad Institute of MIT and Harvard and MIT’s Koch Institute for Integrative Cancer Research. “Without this method, if you want to see things with a high resolution, you have to use very expensive microscopes. What this new technique allows you to do is see things that you couldn’t normally see with standard microscopes. It drives down the cost of imaging because you can see nanoscale things without the need for a specialized facility.”

At the resolution achieved by this technique, which is around 20 nanometers, scientists can see organelles inside cells, as well as clusters of proteins.

“Twenty-fold expansion gets you into the realm that biological molecules operate in. The building blocks of life are nanoscale things: biomolecules, genes, and gene products,” says Edward Boyden, the Y. Eva Tan Professor in Neurotechnology at MIT; a professor of biological engineering, media arts and sciences, and brain and cognitive sciences; a Howard Hughes Medical Institute investigator; and a member of MIT’s McGovern Institute for Brain Research and Koch Institute for Integrative Cancer Research.

Boyden and Kiessling are the senior authors of the new study, which appears today in Nature Methods. MIT graduate student Shiwei Wang and Tay Won Shin PhD ’23 are the lead authors of the paper.

A single expansion

Boyden’s lab invented expansion microscopy in 2015. The technique requires embedding tissue into an absorbent polymer and breaking apart the proteins that normally hold tissue together. When water is added, the gel swells and pulls biomolecules apart from each other.

The original version of this technique, which expanded tissue about fourfold, allowed researchers to obtain images with a resolution of around 70 nanometers. In 2017, Boyden’s lab modified the process to include a second expansion step, achieving an overall 20-fold expansion. This enables even higher resolution, but the process is more complicated.

“We’ve developed several 20-fold expansion technologies in the past, but they require multiple expansion steps,” Boyden says. “If you could do that amount of expansion in a single step, that could simplify things quite a bit.”

With 20-fold expansion, researchers can get down to a resolution of about 20 nanometers, using a conventional light microscope. This allows them see cell structures like microtubules and mitochondria, as well as clusters of proteins.

In the new study, the researchers set out to perform 20-fold expansion with only a single step. This meant that they had to find a gel that was both extremely absorbent and mechanically stable, so that it wouldn’t fall apart when expanded 20-fold.

To achieve that, they used a gel assembled from N,N-dimethylacrylamide (DMAA) and sodium acrylate. Unlike previous expansion gels that rely on adding another molecule to form crosslinks between the polymer strands, this gel forms crosslinks spontaneously and exhibits strong mechanical properties. Such gel components previously had been used in expansion microscopy protocols, but the resulting gels could expand only about tenfold. The MIT team optimized the gel and the polymerization process to make the gel more robust, and to allow for 20-fold expansion.

To further stabilize the gel and enhance its reproducibility, the researchers removed oxygen from the polymer solution prior to gelation, which prevents side reactions that interfere with crosslinking. This step requires running nitrogen gas through the polymer solution, which replaces most of the oxygen in the system.

Once the gel is formed, select bonds in the proteins that hold the tissue together are broken and water is added to make the gel expand. After the expansion is performed, target proteins in tissue can be labeled and imaged.

“This approach may require more sample preparation compared to other super-resolution techniques, but it’s much simpler when it comes to the actual imaging process, especially for 3D imaging,” Shin says. “We document the step-by-step protocol in the manuscript so that readers can go through it easily.”

Imaging tiny structures

Using this technique, the researchers were able to image many tiny structures within brain cells, including structures called synaptic nanocolumns. These are clusters of proteins that are arranged in a specific way at neuronal synapses, allowing neurons to communicate with each other via secretion of neurotransmitters such as dopamine.

In studies of cancer cells, the researchers also imaged microtubules — hollow tubes that help give cells their structure and play important roles in cell division. They were also able to see mitochondria (organelles that generate energy) and even the organization of individual nuclear pore complexes (clusters of proteins that control access to the cell nucleus).

Wang is now using this technique to image carbohydrates known as glycans, which are found on cell surfaces and help control cells’ interactions with their environment. This method could also be used to image tumor cells, allowing scientists to glimpse how proteins are organized within those cells, much more easily than has previously been possible.

The researchers envision that any biology lab should be able to use this technique at a low cost since it relies on standard, off-the-shelf chemicals and common equipment such confocal microscopes and glove bags, which most labs already have or can easily access.

“Our hope is that with this new technology, any conventional biology lab can use this protocol with their existing microscopes, allowing them to approach resolution that can only be achieved with very specialized and costly state-of-the-art microscopes,” Wang says.

The research was funded, in part, by the U.S. National Institutes of Health, an MIT Presidential Graduate Fellowship, U.S. National Science Foundation Graduate Research Fellowship grants, Open Philanthropy, Good Ventures, the Howard Hughes Medical Institute, Lisa Yang, Ashar Aziz, and the European Research Council.

Our brains constantly work to make predictions about what’s going on around us to ensure that we can attend to and consider the unexpected, for instance. A new study examines how this works during consciousness and also breaks down under general anesthesia. The results add evidence to the idea that conscious thought requires synchronized communication — mediated by brain rhythms in specific frequency bands — between basic sensory and higher-order cognitive regions of the brain.

Previously, members of the research team in The Picower Institute for Learning and Memory at MIT and at Vanderbilt University had described how brain rhythms enable the brain to remain prepared to attend to surprises. Cognition-oriented brain regions (generally at the front of the brain) use relatively low-frequency alpha and beta rhythms to suppress processing by sensory regions (generally toward the back of the brain) of stimuli that have become familiar and mundane in the environment (e.g., your co-worker’s music). When sensory regions detect a surprise (e.g., the office fire alarm), they use faster-frequency gamma rhythms to tell the higher regions about it, and the higher regions process that at gamma frequencies to decide what to do (e.g., exit the building).

The new results, published Oct. 7 in the Proceedings of the National Academy of Sciences, show that when animals were under propofol-induced general anesthesia, a sensory region retained the capacity to detect simple surprises but communication with a higher cognitive region toward the front of the brain was lost, making that region unable to engage in its “top-down” regulation of the activity of the sensory region and keeping it oblivious to simple and more complex surprises alike.

What we've got here is failure to communicate

“What we are doing here speaks to the nature of consciousness,” says co-senior author Earl K. Miller, Picower Professor in The Picower Institute for Learning and Memory and MIT’s Department of Brain and Cognitive Sciences. “Propofol general anesthesia deactivates the top-down processes that that underlie cognition. It essentially disconnects communication between the front and back halves of the brain.”

Co-senior author Andre Bastos, an assistant professor in the psychology department at Vanderbilt and a former member of Miller’s MIT lab, adds that the study results highlight the key role of frontal areas in consciousness.

“These results are particularly important given the newfound scientific interest in the mechanisms of consciousness, and how consciousness relates to the ability of the brain to form predictions,” Bastos says.

The brain’s ability to predict is dramatically altered during anesthesia. It was interesting that the front of the brain, areas associated with cognition, were more strongly diminished in their predictive abilities than sensory areas. This suggests that prefrontal areas help to spark an “ignition” event that allows sensory information to become conscious. Sensory cortex activation by itself does not lead to conscious perception. These observations help us narrow down possible models for the mechanisms of consciousness.

Yihan Sophy Xiong, a graduate student in Bastos’ lab who led the study, says the anesthetic reduces the times in which inter-regional communication within the cortex can occur.

“In the awake brain, brain waves give short windows of opportunity for neurons to fire optimally — the ‘refresh rate’ of the brain, so to speak,” Xiong says. “This refresh rate helps organize different brain areas to communicate effectively. Anesthesia both slows down the refresh rate, which narrows these time windows for brain areas to talk to each other and makes the refresh rate less effective, so that neurons become more disorganized about when they can fire. When the refresh rate no longer works as intended, our ability to make predictions is weakened.”

Learning from oddballs

To conduct the research, the neuroscientists measured the electrical signals, “or spiking,” of hundreds of individual neurons and the coordinated rhythms of their aggregated activity (at alpha/beta and gamma frequencies), in two areas on the surface, or cortex, of the brain of two animals as they listened to sequences of tones. Sometimes the sequences would all be the same note (e.g., AAAAA). Sometimes there’d be a simple surprise that the researchers called a “local oddball” (e.g., AAAAB). But sometimes the surprise would be more complicated, or a “global oddball.” For example, after seeing a series of AAAABs, there’d all of a sudden be AAAAA, which violates the global but not the local pattern.

Prior work has suggested that a sensory region (in this case the temporoparietal area, or Tpt) can spot local oddballs on its own, Miller says. Detecting the more complicated global oddball requires the participation of a higher order region (in this case the frontal eye fields, or FEF).

The animals heard the tone sequences both while awake and while under propofol anesthesia. There were no surprises about the waking state. The researchers reaffirmed that top-down alpha/beta rhythms from FEF carried predictions to the Tpt and that Tpt would increase gamma rhythms when an oddball came up, causing FEF (and the prefrontal cortex) to respond with upticks of gamma activity as well.

But by several measures and analyses, the scientists could see these dynamics break down after the animals lost consciousness.

Under propofol, for instance, spiking activity declined overall but when a local oddball came along, Tpt spiking still increased notably but now spiking in FEF didn’t follow suit as it does during wakefulness.

Meanwhile, when a global oddball was presented during wakefulness, the researchers could use software to “decode” representation of that among neurons in FEF and the prefrontal cortex (another cognition-oriented region). They could also decode local oddballs in the Tpt. But under anesthesia the decoder could no longer reliably detect representation of local or global oddballs in FEF or the prefrontal cortex.

Moreover, when they compared rhythms in the regions amid wakeful versus unconscious states they found stark differences. When the animals were awake, oddballs increased gamma activity in both Tpt and FEF and alpha/beta rhythms decreased. Regular, non-oddball stimulation increased alpha/beta rhythms. But when the animals lost consciousness the increase in gamma rhythms from a local oddball was even greater in Tpt than when the animal was awake.

“Under propofol-mediated loss of consciousness, the inhibitory function of alpha/beta became diminished and/or eliminated, leading to disinhibition of oddballs in sensory cortex,” the authors wrote.

Other analyses of inter-region connectivity and synchrony revealed that the regions lost the ability to communicate during anesthesia.

In all, the study’s evidence suggests that conscious thought requires coordination across the cortex, from front to back, the researchers wrote.

“Our results therefore suggest an important role for prefrontal cortex activation, in addition to sensory cortex activation, for conscious perception,” the researchers wrote.

In addition to Xiong, Miller, and Bastos, the paper’s other authors are Jacob Donoghue, Mikael Lundqvist, Meredith Mahnke, Alex Major, and Emery N. Brown.

The National Institutes of Health, The JPB Foundation, and The Picower Institute for Learning and Memory funded the study.

Multimaterial 3D printing enables makers to fabricate customized devices with multiple colors and varied textures. But the process can be time-consuming and wasteful because existing 3D printers must switch between multiple nozzles, often discarding one material before they can start depositing another.

Researchers from MIT and Delft University of Technology have now introduced a more efficient, less wasteful, and higher-precision technique that leverages heat-responsive materials to print objects that have multiple colors, shades, and textures in one step.

Their method, called speed-modulated ironing, utilizes a dual-nozzle 3D printer. The first nozzle deposits a heat-responsive filament and the second nozzle passes over the printed material to activate certain responses, such as changes in opacity or coarseness, using heat.

By controlling the speed of the second nozzle, the researchers can heat the material to specific temperatures, finely tuning the color, shade, and roughness of the heat-responsive filaments. Importantly, this method does not require any hardware modifications.

The researchers developed a model that predicts the amount of heat the “ironing” nozzle will transfer to the material based on its speed. They used this model as the foundation for a user interface that automatically generates printing instructions which achieve color, shade, and texture specifications.

One could use speed-modulated ironing to create artistic effects by varying the color on a printed object. The technique could also produce textured handles that would be easier to grasp for individuals with weakness in their hands.

“Today, we have desktop printers that use a smart combination of a few inks to generate a range of shades and textures. We want to be able to do the same thing with a 3D printer — use a limited set of materials to create a much more diverse set of characteristics for 3D-printed objects,” says Mustafa Doğa Doğan PhD ’24, co-author of a paper on speed-modulated ironing.

This project is a collaboration between the research groups of Zjenja Doubrovski, assistant professor at TU Delft, and Stefanie Mueller, the TIBCO Career Development Professor in the Department of Electrical Engineering and Computer Science (EECS) at MIT and a member of the MIT Computer Science and Artificial Intelligence Laboratory (CSAIL). Doğan worked closely with lead author Mehmet Ozdemir of TU Delft; Marwa AlAlawi, a mechanical engineering graduate student at MIT; and Jose Martinez Castro of TU Delft. The research will be presented at the ACM Symposium on User Interface Software and Technology.

Modulating speed to control temperature

The researchers launched the project to explore better ways to achieve multiproperty 3D printing with a single material. The use of heat-responsive filaments was promising, but most existing methods use a single nozzle to do printing and heating. The printer always needs to first heat the nozzle to the desired target temperature before depositing the material.

However, heating and cooling the nozzle takes a long time, and there is a danger that the filament in the nozzle might degrade as it reaches higher temperatures.

To prevent these problems, the team developed an ironing technique where material is printed using one nozzle, then activated by a second, empty nozzle which only reheats it. Instead of adjusting the temperature to trigger the material response, the researchers keep the temperature of the second nozzle constant and vary the speed at which it moves over the printed material, slightly touching the top of the layer.
 

“As we modulate the speed, that allows the printed layer we are ironing to reach different temperatures. It is similar to what happens if you move your finger over a flame. If you move it quickly, you might not be burned, but if you drag it across the flame slowly, your finger will reach a higher temperature,” AlAlawi says.

The MIT team collaborated with the TU Delft researchers to develop the theoretical model that predicts how fast the second nozzle must move to heat the material to a specific temperature.

The model correlates a material’s output temperature with its heat-responsive properties to determine the exact nozzle speed which will achieve certain colors, shades, or textures in the printed object.

“There are a lot of inputs that can affect the results we get. We are modeling something that is very complicated, but we also want to make sure the results are fine-grained,” AlAlawi says.

The team dug into scientific literature to determine proper heat transfer coefficients for a set of unique materials, which they built into their model. They also had to contend with an array of unpredictable variables, such as heat that may be dissipated by fans and the air temperature in the room where the object is being printed.

They incorporated the model into a user-friendly interface that simplifies the scientific process, automatically translating the pixels in a maker’s 3D model into a set of machine instructions that control the speed at which the object is printed and ironed by the dual nozzles.

Faster, finer fabrication

They tested their approach with three heat-responsive filaments. The first, a foaming polymer with particles that expand as they are heated, yields different shades, translucencies, and textures. They also experimented with a filament filled with wood fibers and one with cork fibers, both of which can be charred to produce increasingly darker shades.

The researchers demonstrated how their method could produce objects like water bottles that are partially translucent. To make the water bottles, they ironed the foaming polymer at low speeds to create opaque regions and higher speeds to create translucent ones. They also utilized the foaming polymer to fabricate a bike handle with varied roughness to improve a rider’s grip.

Trying to produce similar objects using traditional multimaterial 3D printing took far more time, sometimes adding hours to the printing process, and consumed more energy and material. In addition, speed-modulated ironing could produce fine-grained shade and texture gradients that other methods could not achieve.

In the future, the researchers want to experiment with other thermally responsive materials, such as plastics. They also hope to explore the use of speed-modulated ironing to modify the mechanical and acoustic properties of certain materials.

A growing portion of Americans who are struggling to pay for their household energy live in the South and Southwest, reflecting a climate-driven shift away from heating needs and toward air conditioning use, an MIT study finds.

The newly published research also reveals that a major U.S. federal program that provides energy subsidies to households, by assigning block grants to states, does not yet fully match these recent trends.

The work evaluates the “energy burden” on households, which reflects the percentage of income needed to pay for energy necessities, from 2015 to 2020. Households with an energy burden greater than 6 percent of income are considered to be in “energy poverty.” With climate change, rising temperatures are expected to add financial stress in the South, where air conditioning is increasingly needed. Meanwhile, milder winters are expected to reduce heating costs in some colder regions.

“From 2015 to 2020, there is an increase in burden generally, and you do also see this southern shift,” says Christopher Knittel, an MIT energy economist and co-author of a new paper detailing the study’s results. About federal aid, he adds, “When you compare the distribution of the energy burden to where the money is going, it’s not aligned too well.”

The paper, “U.S. federal resource allocations are inconsistent with concentrations of energy poverty,” is published today in Science Advances.

The authors are Carlos Batlle, a professor at Comillas University in Spain and a senior lecturer with the MIT Energy Initiative; Peter Heller SM ’24, a recent graduate of the MIT Technology and Policy Program; Knittel, the George P. Shultz Professor at the MIT Sloan School of Management and associate dean for climate and sustainability at MIT; and Tim Schittekatte, a senior lecturer at MIT Sloan.

A scorching decade

The study, which grew out of graduate research that Heller conducted at MIT, deploys a machine-learning estimation technique that the scholars applied to U.S. energy use data.

Specifically, the researchers took a sample of about 20,000 households from the U.S. Energy Information Administration’s Residential Energy Consumption Survey, which includes a wide variety of demographic characteristics about residents, along with building-type and geographic information. Then, using the U.S. Census Bureau’s American Community Survey data for 2015 and 2020, the research team estimated the average household energy burden for every census tract in the lower 48 states — 73,057 in 2015, and 84,414 in 2020.

That allowed the researchers to chart the changes in energy burden in recent years, including the shift toward a greater energy burden in southern states. In 2015, Maine, Mississippi, Arkansas, Vermont, and Alabama were the five states (ranked in descending order) with the highest energy burden across census bureau tracts. In 2020, that had shifted somewhat, with Maine and Vermont dropping on the list and southern states increasingly having a larger energy burden. That year, the top five states in descending order were Mississippi, Arkansas, Alabama, West Virginia, and Maine.

The data also reflect a urban-rural shift. In 2015, 23 percent of the census tracts where the average household is living in energy poverty were urban. That figure shrank to 14 percent by 2020.

All told, the data are consistent with the picture of a warming world, in which milder winters in the North, Northwest, and Mountain West require less heating fuel, while more extreme summer temperatures in the South require more air conditioning.

“Who’s going to be harmed most from climate change?” asks Knittel. “In the U.S., not surprisingly, it’s going to be the southern part of the U.S. And our study is confirming that, but also suggesting it’s the southern part of the U.S that’s least able to respond. If you’re already burdened, the burden’s growing.”

An evolution for LIHEAP?

In addition to identifying the shift in energy needs during the last decade, the study also illuminates a longer-term change in U.S. household energy needs, dating back to the 1980s. The researchers compared the present-day geography of U.S. energy burden to the help currently provided by the federal Low Income Home Energy Assistance Program (LIHEAP), which dates to 1981.

Federal aid for energy needs actually predates LIHEAP, but the current program was introduced in 1981, then updated in 1984 to include cooling needs such as air conditioning. When the formula was updated in 1984, two “hold harmless” clauses were also adopted, guaranteeing states a minimum amount of funding.

Still, LIHEAP’s parameters also predate the rise of temperatures over the last 40 years, and the current study shows that, compared to the current landscape of energy poverty, LIHEAP distributes relatively less of its funding to southern and southwestern states.

“The way Congress uses formulas set in the 1980s keeps funding distributions nearly the same as it was in the 1980s,” Heller observes. “Our paper illustrates the shift in need that has occurred over the decades since then.”

Currently, it would take a fourfold increase in LIHEAP to ensure that no U.S. household experiences energy poverty. But the researchers tested out a new funding design, which would help the worst-off households first, nationally, ensuring that no household would have an energy burden of greater than 20.3 percent.

“We think that’s probably the most equitable way to allocate the money, and by doing that, you now have a different amount of money that should go to each state, so that no one state is worse off than the others,” Knittel says.

And while the new distribution concept would require a certain amount of subsidy reallocation among states, it would be with the goal of helping all households avoid a certain level of energy poverty, across the country, at a time of changing climate, warming weather, and shifting energy needs in the U.S.

“We can optimize where we spend the money, and that optimization approach is an important thing to think about,” Knittel says. 

A recent award from the U.S. Defense Advanced Research Projects Agency (DARPA) brings together researchers from Massachusetts Institute of Technology (MIT), Carnegie Mellon University (CMU), and Lehigh University (Lehigh) under the Multiobjective Engineering and Testing of Alloy Structures (METALS) program. The team will research novel design tools for the simultaneous optimization of shape and compositional gradients in multi-material structures that complement new high-throughput materials testing techniques, with particular attention paid to the bladed disk (blisk) geometry commonly found in turbomachinery (including jet and rocket engines) as an exemplary challenge problem.

“This project could have important implications across a wide range of aerospace technologies. Insights from this work may enable more reliable, reusable, rocket engines that will power the next generation of heavy-lift launch vehicles,” says Zachary Cordero, the Esther and Harold E. Edgerton Associate Professor in the MIT Department of Aeronautics and Astronautics (AeroAstro) and the project’s lead principal investigator. “This project merges classical mechanics analyses with cutting-edge generative AI design technologies to unlock the plastic reserve of compositionally graded alloys allowing safe operation in previously inaccessible conditions.”

Different locations in blisks require different thermomechanical properties and performance, such as resistance to creep, low cycle fatigue, high strength, etc. Large scale production also necessitates consideration of cost and sustainability metrics such as sourcing and recycling of alloys in the design.

“Currently, with standard manufacturing and design procedures, one must come up with a single magical material, composition, and processing parameters to meet ‘one part-one material’ constraints,” says Cordero. “Desired properties are also often mutually exclusive prompting inefficient design tradeoffs and compromises.”

Although a one-material approach may be optimal for a singular location in a component, it may leave other locations exposed to failure or may require a critical material to be carried throughout an entire part when it may only be needed in a specific location. With the rapid advancement of additive manufacturing processes that are enabling voxel-based composition and property control, the team sees unique opportunities for leap-ahead performance in structural components are now possible.

Cordero’s collaborators include Zoltan Spakovszky, the T. Wilson (1953) Professor in Aeronautics in AeroAstro; A. John Hart, the Class of 1922 Professor and head of the Department of Mechanical Engineering; Faez Ahmed, ABS Career Development Assistant Professor of mechanical engineering at MIT; S. Mohadeseh Taheri-Mousavi, assistant professor of materials science and engineering at CMU; and Natasha Vermaak, associate professor of mechanical engineering and mechanics at Lehigh.

The team’s expertise spans hybrid integrated computational material engineering and machine-learning-based material and process design, precision instrumentation, metrology, topology optimization, deep generative modeling, additive manufacturing, materials characterization, thermostructural analysis, and turbomachinery.

“It is especially rewarding to work with the graduate students and postdoctoral researchers collaborating on the METALS project, spanning from developing new computational approaches to building test rigs operating under extreme conditions,” says Hart. “It is a truly unique opportunity to build breakthrough capabilities that could underlie propulsion systems of the future, leveraging digital design and manufacturing technologies.”

This research is funded by DARPA under contract HR00112420303. The views, opinions, and/or findings expressed are those of the author and should not be interpreted as representing the official views or policies of the Department of Defense or the U.S. government and no official endorsement should be inferred.

MIT researchers have some good environmental news: Mercury emissions from human activity have been declining over the past two decades, despite global emissions inventories that indicate otherwise.

In a new study, the researchers analyzed measurements from all available monitoring stations in the Northern Hemisphere and found that atmospheric concentrations of mercury declined by about 10 percent between 2005 and 2020.

They used two separate modeling methods to determine what is driving that trend. Both techniques pointed to a decline in mercury emissions from human activity as the most likely cause.

Global inventories, on the other hand, have reported opposite trends. These inventories estimate atmospheric emissions using models that incorporate average emission rates of polluting activities and the scale of these activities worldwide.

“Our work shows that it is very important to learn from actual, on-the-ground data to try and improve our models and these emissions estimates. This is very relevant for policy because, if we are not able to accurately estimate past mercury emissions, how are we going to predict how mercury pollution will evolve in the future?” says Ari Feinberg, a former postdoc in the Institute for Data, Systems, and Society (IDSS) and lead author of the study.

The new results could help inform scientists who are embarking on a collaborative, global effort to evaluate pollution models and develop a more in-depth understanding of what drives global atmospheric concentrations of mercury.

However, due to a lack of data from global monitoring stations and limitations in the scientific understanding of mercury pollution, the researchers couldn’t pinpoint a definitive reason for the mismatch between the inventories and the recorded measurements.

“It seems like mercury emissions are moving in the right direction, and could continue to do so, which is heartening to see. But this was as far as we could get with mercury. We need to keep measuring and advancing the science,” adds co-author Noelle Selin, an MIT professor in the IDSS and the Department of Earth, Atmospheric and Planetary Sciences (EAPS).

Feinberg and Selin, his MIT postdoctoral advisor, are joined on the paper by an international team of researchers that contributed atmospheric mercury measurement data and statistical methods to the study. The research appears this week in the Proceedings of the National Academy of Sciences.

Mercury mismatch

The Minamata Convention is a global treaty that aims to cut human-caused emissions of mercury, a potent neurotoxin that enters the atmosphere from sources like coal-fired power plants and small-scale gold mining.

The treaty, which was signed in 2013 and went into force in 2017, is evaluated every five years. The first meeting of its conference of parties coincided with disheartening news reports that said global inventories of mercury emissions, compiled in part from information from national inventories, had increased despite international efforts to reduce them.

This was puzzling news for environmental scientists like Selin. Data from monitoring stations showed atmospheric mercury concentrations declining during the same period.

Bottom-up inventories combine emission factors, such as the amount of mercury that enters the atmosphere when coal mined in a certain region is burned, with estimates of pollution-causing activities, like how much of that coal is burned in power plants.

“The big question we wanted to answer was: What is actually happening to mercury in the atmosphere and what does that say about anthropogenic emissions over time?” Selin says.

Modeling mercury emissions is especially tricky. First, mercury is the only metal that is in liquid form at room temperature, so it has unique properties. Moreover, mercury that has been removed from the atmosphere by sinks like the ocean or land can be re-emitted later, making it hard to identify primary emission sources.

At the same time, mercury is more difficult to study in laboratory settings than many other air pollutants, especially due to its toxicity, so scientists have limited understanding of all chemical reactions mercury can undergo. There is also a much smaller network of mercury monitoring stations, compared to other polluting gases like methane and nitrous oxide.

“One of the challenges of our study was to come up with statistical methods that can address those data gaps, because available measurements come from different time periods and different measurement networks,” Feinberg says.

Multifaceted models

The researchers compiled data from 51 stations in the Northern Hemisphere. They used statistical techniques to aggregate data from nearby stations, which helped them overcome data gaps and evaluate regional trends.

By combining data from 11 regions, their analysis indicated that Northern Hemisphere atmospheric mercury concentrations declined by about 10 percent between 2005 and 2020.

Then the researchers used two modeling methods — biogeochemical box modeling and chemical transport modeling — to explore possible causes of that decline.  Box modeling was used to run hundreds of thousands of simulations to evaluate a wide array of emission scenarios. Chemical transport modeling is more computationally expensive but enables researchers to assess the impacts of meteorology and spatial variations on trends in selected scenarios.

For instance, they tested one hypothesis that there may be an additional environmental sink that is removing more mercury from the atmosphere than previously thought. The models would indicate the feasibility of an unknown sink of that magnitude.

“As we went through each hypothesis systematically, we were pretty surprised that we could really point to declines in anthropogenic emissions as being the most likely cause,” Selin says.

Their work underscores the importance of long-term mercury monitoring stations, Feinberg adds. Many stations the researchers evaluated are no longer operational because of a lack of funding.

While their analysis couldn’t zero in on exactly why the emissions inventories didn’t match up with actual data, they have a few hypotheses.

One possibility is that global inventories are missing key information from certain countries. For instance, the researchers resolved some discrepancies when they used a more detailed regional inventory from China. But there was still a gap between observations and estimates.

They also suspect the discrepancy might be the result of changes in two large sources of mercury that are particularly uncertain: emissions from small-scale gold mining and mercury-containing products.

Small-scale gold mining involves using mercury to extract gold from soil and is often performed in remote parts of developing countries, making it hard to estimate. Yet small-scale gold mining contributes about 40 percent of human-made emissions.

In addition, it’s difficult to determine how long it takes the pollutant to be released into the atmosphere from discarded products like thermometers or scientific equipment.

“We’re not there yet where we can really pinpoint which source is responsible for this discrepancy,” Feinberg says.

In the future, researchers from multiple countries, including MIT, will collaborate to study and improve the models they use to estimate and evaluate emissions. This research will be influential in helping that project move the needle on monitoring mercury, he says.

This research was funded by the Swiss National Science Foundation, the U.S. National Science Foundation, and the U.S. Environmental Protection Agency.

Industrial electrochemical processes that use electrodes to produce fuels and chemical products are hampered by the formation of bubbles that block parts of the electrode surface, reducing the area available for the active reaction. Such blockage reduces the performance of the electrodes by anywhere from 10 to 25 percent.

But new research reveals a decades-long misunderstanding about the extent of that interference. The findings show exactly how the blocking effect works and could lead to new ways of designing electrode surfaces to minimize inefficiencies in these widely used electrochemical processes.

It has long been assumed that the entire area of the electrode shadowed by each bubble would be effectively inactivated. But it turns out that a much smaller area — roughly the area where the bubble actually contacts the surface — is blocked from its electrochemical activity. The new insights could lead directly to new ways of patterning the surfaces to minimize the contact area and improve overall efficiency.

The findings are reported today in the journal Nanoscale, in a paper by recent MIT graduate Jack Lake PhD ’23, graduate student Simon Rufer, professor of mechanical engineering Kripa Varanasi, research scientist Ben Blaiszik, and six others at the University of Chicago and Argonne National Laboratory. The team has made available an open-source, AI-based software tool that engineers and scientists can now use to automatically recognize and quantify bubbles formed on a given surface, as a first step toward controlling the electrode material’s properties.

Gas-evolving electrodes, often with catalytic surfaces that promote chemical reactions, are used in a wide variety of processes, including the production of “green” hydrogen without the use of fossil fuels, carbon-capture processes that can reduce greenhouse gas emissions, aluminum production, and the chlor-alkali process that is used to make widely used chemical products.

These are very widespread processes. The chlor-alkali process alone accounts for 2 percent of all U.S. electricity usage; aluminum production accounts for 3 percent of global electricity; and both carbon capture and hydrogen production are likely to grow rapidly in coming years as the world strives to meet greenhouse-gas reduction targets. So, the new findings could make a real difference, Varanasi says.

“Our work demonstrates that engineering the contact and growth of bubbles on electrodes can have dramatic effects” on how bubbles form and how they leave the surface, he says. “The knowledge that the area under bubbles can be significantly active ushers in a new set of design rules for high-performance electrodes to avoid the deleterious effects of bubbles.”

“The broader literature built over the last couple of decades has suggested that not only that small area of contact but the entire area under the bubble is passivated,” Rufer says. The new study reveals “a significant difference between the two models because it changes how you would develop and design an electrode to minimize these losses.”

To test and demonstrate the implications of this effect, the team produced different versions of electrode surfaces with patterns of dots that nucleated and trapped bubbles at different sizes and spacings. They were able to show that surfaces with widely spaced dots promoted large bubble sizes but only tiny areas of surface contact, which helped to make clear the difference between the expected and actual effects of bubble coverage.

Developing the software to detect and quantify bubble formation was necessary for the team’s analysis, Rufer explains. “We wanted to collect a lot of data and look at a lot of different electrodes and different reactions and different bubbles, and they all look slightly different,” he says. Creating a program that could deal with different materials and different lighting and reliably identify and track the bubbles was a tricky process, and machine learning was key to making it work, he says.

Using that tool, he says, they were able to collect “really significant amounts of data about the bubbles on a surface, where they are, how big they are, how fast they’re growing, all these different things.” The tool is now freely available for anyone to use via the GitHub repository.

By using that tool to correlate the visual measures of bubble formation and evolution with electrical measurements of the electrode’s performance, the researchers were able to disprove the accepted theory and to show that only the area of direct contact is affected. Videos further proved the point, revealing new bubbles actively evolving directly under parts of a larger bubble.

The researchers developed a very general methodology that can be applied to characterize and understand the impact of bubbles on any electrode or catalyst surface. They were able to quantify the bubble passivation effects in a new performance metric they call BECSA (Bubble-induced electrochemically active surface), as opposed to ECSA (electrochemically active surface area), that is used in the field. “The BECSA metric was a concept we defined in an earlier study but did not have an effective method to estimate until this work,” says Varanasi.

The knowledge that the area under bubbles can be significantly active ushers in a new set of design rules for high-performance electrodes. This means that electrode designers should seek to minimize bubble contact area rather than simply bubble coverage, which can be achieved by controlling the morphology and chemistry of the electrodes. Surfaces engineered to control bubbles can not only improve the overall efficiency of the processes and thus reduce energy use, they can also save on upfront materials costs. Many of these gas-evolving electrodes are coated with catalysts made of expensive metals like platinum or iridium, and the findings from this work can be used to engineer electrodes to reduce material wasted by reaction-blocking bubbles.

Varanasi says that “the insights from this work could inspire new electrode architectures that not only reduce the usage of precious materials, but also improve the overall electrolyzer performance,” both of which would provide large-scale environmental benefits.

The research team included Jim James, Nathan Pruyne, Aristana Scourtas, Marcus Schwarting, Aadit Ambalkar, Ian Foster, and Ben Blaiszik at the University of Chicago and Argonne National Laboratory. The work was supported by the U.S. Department of Energy under the ARPA-E program. This work made use of the MIT.nano facilities.

MIT engineers have built a new desalination system that runs with the rhythms of the sun.

The solar-powered system removes salt from water at a pace that closely follows changes in solar energy. As sunlight increases through the day, the system ramps up its desalting process and automatically adjusts to any sudden variation in sunlight, for example by dialing down in response to a passing cloud or revving up as the skies clear.

Because the system can quickly react to subtle changes in sunlight, it maximizes the utility of solar energy, producing large quantities of clean water despite variations in sunlight throughout the day. In contrast to other solar-driven desalination designs, the MIT system requires no extra batteries for energy storage, nor a supplemental power supply, such as from the grid.

The engineers tested a community-scale prototype on groundwater wells in New Mexico over six months, working in variable weather conditions and water types. The system harnessed on average over 94 percent of the electrical energy generated from the system’s solar panels to produce up to 5,000 liters of water per day despite large swings in weather and available sunlight.

“Conventional desalination technologies require steady power and need battery storage to smooth out a variable power source like solar. By continually varying power consumption in sync with the sun, our technology directly and efficiently uses solar power to make water,” says Amos Winter, the Germeshausen Professor of Mechanical Engineering and director of the K. Lisa Yang Global Engineering and Research (GEAR) Center at MIT. “Being able to make drinking water with renewables, without requiring battery storage, is a massive grand challenge. And we’ve done it.”

The system is geared toward desalinating brackish groundwater — a salty source of water that is found in underground reservoirs and is more prevalent than fresh groundwater resources. The researchers see brackish groundwater as a huge untapped source of potential drinking water, particularly as reserves of fresh water are stressed in parts of the world. They envision that the new renewable, battery-free system could provide much-needed drinking water at low costs, especially for inland communities where access to seawater and grid power are limited.

“The majority of the population actually lives far enough from the coast, that seawater desalination could never reach them. They consequently rely heavily on groundwater, especially in remote, low-income regions. And unfortunately, this groundwater is becoming more and more saline due to climate change,” says Jonathan Bessette, MIT PhD student in mechanical engineering. “This technology could bring sustainable, affordable clean water to underreached places around the world.”

The researchers report details the new system in a paper appearing today in Nature Water. The study’s co-authors are Bessette, Winter, and staff engineer Shane Pratt.

Pump and flow

The new system builds on a previous design, which Winter and his colleagues, including former MIT postdoc Wei He, reported earlier this year. That system aimed to desalinate water through “flexible batch electrodialysis.”

Electrodialysis and reverse osmosis are two of the main methods used to desalinate brackish groundwater. With reverse osmosis, pressure is used to pump salty water through a membrane and filter out salts. Electrodialysis uses an electric field to draw out salt ions as water is pumped through a stack of ion-exchange membranes.

Scientists have looked to power both methods with renewable sources. But this has been especially challenging for reverse osmosis systems, which traditionally run at a steady power level that’s incompatible with naturally variable energy sources such as the sun.

Winter, He, and their colleagues focused on electrodialysis, seeking ways to make a more flexible, “time-variant” system that would be responsive to variations in renewable, solar power.

In their previous design, the team built an electrodialysis system consisting of water pumps, an ion-exchange membrane stack, and a solar panel array. The innovation in this system was a model-based control system that used sensor readings from every part of the system to predict the optimal rate at which to pump water through the stack and the voltage that should be applied to the stack to maximize the amount of salt drawn out of the water.

When the team tested this system in the field, it was able to vary its water production with the sun’s natural variations. On average, the system directly used 77 percent of the available electrical energy produced by the solar panels, which the team estimated was 91 percent more than traditionally designed solar-powered electrodialysis systems.

Still, the researchers felt they could do better.

“We could only calculate every three minutes, and in that time, a cloud could literally come by and block the sun,” Winter says. “The system could be saying, ‘I need to run at this high power.’ But some of that power has suddenly dropped because there’s now less sunlight. So, we had to make up that power with extra batteries.”

Solar commands

In their latest work, the researchers looked to eliminate the need for batteries, by shaving the system’s response time to a fraction of a second. The new system is able to update its desalination rate, three to five times per second. The faster response time enables the system to adjust to changes in sunlight throughout the day, without having to make up any lag in power with additional power supplies.

The key to the nimbler desalting is a simpler control strategy, devised by Bessette and Pratt. The new strategy is one of “flow-commanded current control,” in which the system first senses the amount of solar power that is being produced by the system’s solar panels. If the panels are generating more power than the system is using, the controller automatically “commands” the system to dial up its pumping, pushing more water through the electrodialysis stacks. Simultaneously, the system diverts some of the additional solar power by increasing the electrical current delivered to the stack, to drive more salt out of the faster-flowing water.

“Let’s say the sun is rising every few seconds,” Winter explains. “So, three times a second, we’re looking at the solar panels and saying, ‘Oh, we have more power — let’s bump up our flow rate and current a little bit.’ When we look again and see there’s still more excess power, we’ll up it again. As we do that, we’re able to closely match our consumed power with available solar power really accurately, throughout the day. And the quicker we loop this, the less battery buffering we need.”

The engineers incorporated the new control strategy into a fully automated system that they sized to desalinate brackish groundwater at a daily volume that would be enough to supply a small community of about 3,000 people. They operated the system for six months on several wells at the Brackish Groundwater National Desalination Research Facility in Alamogordo, New Mexico. Throughout the trial, the prototype operated under a wide range of solar conditions, harnessing over 94 percent of the solar panel’s electrical energy, on average, to directly power desalination.

“Compared to how you would traditionally design a solar desal system, we cut our required battery capacity by almost 100 percent,” Winter says.

The engineers plan to further test and scale up the system in hopes of supplying larger communities, and even whole municipalities, with low-cost, fully sun-driven drinking water.

“While this is a major step forward, we’re still working diligently to continue developing lower cost, more sustainable desalination methods,” Bessette says.

“Our focus now is on testing, maximizing reliability, and building out a product line that can provide desalinated water using renewables to multiple markets around the world," Pratt adds.

The team will be launching a company based on their technology in the coming months.

This research was supported in part by the National Science Foundation, the Julia Burke Foundation, and the MIT Morningside Academy of Design. This work was additionally supported in-kind by Veolia Water Technologies and Solutions and Xylem Goulds. 

Since the 1950s, a chemotherapy drug known as 5-fluorouracil has been used to treat many types of cancer, including blood cancers and cancers of the digestive tract.

Doctors have long believed that this drug works by damaging the building blocks of DNA. However, a new study from MIT has found that in cancers of the colon and other gastrointestinal cancers, it actually kills cells by interfering with RNA synthesis.

The findings could have a significant effect on how doctors treat many cancer patients. Usually, 5-fluorouracil is given in combination with chemotherapy drugs that damage DNA, but the new study found that for colon cancer, this combination does not achieve the synergistic effects that were hoped for. Instead, combining 5-FU with drugs that affect RNA synthesis could make it more effective in patients with GI cancers, the researchers say.

“Our work is the most definitive study to date showing that RNA incorporation of the drug, leading to an RNA damage response, is responsible for how the drug works in GI cancers,” says Michael Yaffe, a David H. Koch Professor of Science at MIT, the director of the MIT Center for Precision Cancer Medicine, and a member of MIT’s Koch Institute for Integrative Cancer Research. “Textbooks implicate the DNA effects of the drug as the mechanism in all cancer types, but our data shows that RNA damage is what’s really important for the types of tumors, like GI cancers, where the drug is used clinically.”

Yaffe, the senior author of the new study, hopes to plan clinical trials of 5-fluorouracil with drugs that would enhance its RNA-damaging effects and kill cancer cells more effectively.

Jung-Kuei Chen, a Koch Institute research scientist, and Karl Merrick, a former MIT postdoc, are the lead authors of the paper, which appears today in Cell Reports Medicine.

An unexpected mechanism

Clinicians use 5-fluorouracil (5-FU) as a first-line drug for colon, rectal, and pancreatic cancers. It’s usually given in combination with oxaliplatin or irinotecan, which damage DNA in cancer cells. The combination was thought to be effective because 5-FU can disrupt the synthesis of DNA nucleotides. Without those building blocks, cells with damaged DNA wouldn’t be able to efficiently repair the damage and would undergo cell death.

Yaffe’s lab, which studies cell signaling pathways, wanted to further explore the underlying mechanisms of how these drug combinations preferentially kill cancer cells.

The researchers began by testing 5-FU in combination with oxaliplatin or irinotecan in colon cancer cells grown in the lab. To their surprise, they found that not only were the drugs not synergistic, in many cases they were less effective at killing cancer cells than what one would expect by simply adding together the effects of 5-FU or the DNA-damaging drug given alone.

“One would have expected that these combinations to cause synergistic cancer cell death because you are targeting two different aspects of a shared process: breaking DNA, and making nucleotides,” Yaffe says. “Karl looked at a dozen colon cancer cell lines, and not only were the drugs not synergistic, in most cases they were antagonistic. One drug seemed to be undoing what the other drug was doing.”

Yaffe’s lab then teamed up with Adam Palmer, an assistant professor of pharmacology at the University of North Carolina School of Medicine, who specializes in analyzing data from clinical trials. Palmer’s research group examined data from colon cancer patients who had been on one or more of these drugs and showed that the drugs did not show synergistic effects on survival in most patients.

“This confirmed that when you give these combinations to people, it’s not generally true that the drugs are actually working together in a beneficial way within an individual patient,” Yaffe says. “Instead, it appears that one drug in the combination works well for some patients while another drug in the combination works well in other patients. We just cannot yet predict which drug by itself is best for which patient, so everyone gets the combination.”

These results led the researchers to wonder just how 5-FU was working, if not by disrupting DNA repair. Studies in yeast and mammalian cells had shown that the drug also gets incorporated into RNA nucleotides, but there has been dispute over how much this RNA damage contributes to the drug’s toxic effects on cancer cells.

Inside cells, 5-FU is broken down into two different metabolites. One of these gets incorporated into DNA nucleotides, and other into RNA nucleotides. In studies of colon cancer cells, the researchers found that the metabolite that interferes with RNA was much more effective at killing colon cancer cells than the one that disrupts DNA.

That RNA damage appears to primarily affect ribosomal RNA, a molecule that forms part of the ribosome — a cell organelle responsible for assembling new proteins. If cells can’t form new ribosomes, they can’t produce enough proteins to function. Additionally, the lack of undamaged ribosomal RNA causes cells to destroy a large set of proteins that normally bind up the RNA to make new functional ribosomes.

The researchers are now exploring how this ribosomal RNA damage leads cells to under programmed cell death, or apoptosis. They hypothesize that sensing of the damaged RNAs within cell structures called lysosomes somehow triggers an apoptotic signal.

“My lab is very interested in trying to understand the signaling events during disruption of ribosome biogenesis, particularly in GI cancers and even some ovarian cancers, that cause the cells to die. Somehow, they must be monitoring the quality control of new ribosome synthesis, which somehow is connected to the death pathway machinery,” Yaffe says.

New combinations

The findings suggest that drugs that stimulate ribosome production could work together with 5-FU to make a highly synergistic combination. In their study, the researchers showed that a molecule that inhibits KDM2A, a suppressor of ribosome production, helped to boost the rate of cell death in colon cancer cells treated with 5-FU.

The findings also suggest a possible explanation for why combining 5-FU with a DNA-damaging drug often makes both drugs less effective. Some DNA damaging drugs send a signal to the cell to stop making new ribosomes, which would negate 5-FU’s effect on RNA. A better approach may be to give each drug a few days apart, which would give patients the potential benefits of each drug, without having them cancel each other out.

“Importantly, our data doesn’t say that these combination therapies are wrong. We know they’re effective clinically. It just says that if you adjust how you give these drugs, you could potentially make those therapies even better, with relatively minor changes in the timing of when the drugs are given,” Yaffe says.

He is now hoping to work with collaborators at other institutions to run a phase 2 or 3 clinical trial in which patients receive the drugs on an altered schedule.

“A trial is clearly needed to look for efficacy, but it should be straightforward to initiate because these are already clinically accepted drugs that form the standard of care for GI cancers. All we’re doing is changing the timing with which we give them,” he says.

The researchers also hope that their work could lead to the identification of biomarkers that predict which patients’ tumors will be more susceptible to drug combinations that include 5-FU. One such biomarker could be RNA polymerase I, which is active when cells are producing a lot of ribosomal RNA.

The research was funded by the Damon Runyon Cancer Research Foundation, a fellowship from the Ludwig Center at MIT, the National Institutes of Health, the Ovarian Cancer Research Fund, the Charles and Marjorie Holloway Foundation, and the STARR Cancer Consortium.

Imagine you’re tasked with sending a team of football players onto a field to assess the condition of the grass (a likely task for them, of course). If you pick their positions randomly, they might cluster together in some areas while completely neglecting others. But if you give them a strategy, like spreading out uniformly across the field, you might get a far more accurate picture of the grass condition.

Now, imagine needing to spread out not just in two dimensions, but across tens or even hundreds. That's the challenge MIT Computer Science and Artificial Intelligence Laboratory (CSAIL) researchers are getting ahead of. They've developed an AI-driven approach to “low-discrepancy sampling,” a method that improves simulation accuracy by distributing data points more uniformly across space.

A key novelty lies in using graph neural networks (GNNs), which allow points to “communicate” and self-optimize for better uniformity. Their approach marks a pivotal enhancement for simulations in fields like robotics, finance, and computational science, particularly in handling complex, multidimensional problems critical for accurate simulations and numerical computations.

“In many problems, the more uniformly you can spread out points, the more accurately you can simulate complex systems,” says T. Konstantin Rusch, lead author of the new paper and MIT CSAIL postdoc. “We've developed a method called Message-Passing Monte Carlo (MPMC) to generate uniformly spaced points, using geometric deep learning techniques. This further allows us to generate points that emphasize dimensions which are particularly important for a problem at hand, a property that is highly important in many applications. The model’s underlying graph neural networks lets the points 'talk' with each other, achieving far better uniformity than previous methods.”

Their work was published in the September issue of the Proceedings of the National Academy of Sciences.

Take me to Monte Carlo

The idea of Monte Carlo methods is to learn about a system by simulating it with random sampling. Sampling is the selection of a subset of a population to estimate characteristics of the whole population. Historically, it was already used in the 18th century,  when mathematician Pierre-Simon Laplace employed it to estimate the population of France without having to count each individual.

Low-discrepancy sequences, which are sequences with low discrepancy, i.e., high uniformity, such as Sobol’, Halton, and Niederreiter, have long been the gold standard for quasi-random sampling, which exchanges random sampling with low-discrepancy sampling. They are widely used in fields like computer graphics and computational finance, for everything from pricing options to risk assessment, where uniformly filling spaces with points can lead to more accurate results. 

The MPMC framework suggested by the team transforms random samples into points with high uniformity. This is done by processing the random samples with a GNN that minimizes a specific discrepancy measure.

One big challenge of using AI for generating highly uniform points is that the usual way to measure point uniformity is very slow to compute and hard to work with. To solve this, the team switched to a quicker and more flexible uniformity measure called L2-discrepancy. For high-dimensional problems, where this method isn’t enough on its own, they use a novel technique that focuses on important lower-dimensional projections of the points. This way, they can create point sets that are better suited for specific applications.

The implications extend far beyond academia, the team says. In computational finance, for example, simulations rely heavily on the quality of the sampling points. “With these types of methods, random points are often inefficient, but our GNN-generated low-discrepancy points lead to higher precision,” says Rusch. “For instance, we considered a classical problem from computational finance in 32 dimensions, where our MPMC points beat previous state-of-the-art quasi-random sampling methods by a factor of four to 24.”

Robots in Monte Carlo

In robotics, path and motion planning often rely on sampling-based algorithms, which guide robots through real-time decision-making processes. The improved uniformity of MPMC could lead to more efficient robotic navigation and real-time adaptations for things like autonomous driving or drone technology. “In fact, in a recent preprint, we demonstrated that our MPMC points achieve a fourfold improvement over previous low-discrepancy methods when applied to real-world robotics motion planning problems,” says Rusch.

“Traditional low-discrepancy sequences were a major advancement in their time, but the world has become more complex, and the problems we're solving now often exist in 10, 20, or even 100-dimensional spaces,” says Daniela Rus, CSAIL director and MIT professor of electrical engineering and computer science. “We needed something smarter, something that adapts as the dimensionality grows. GNNs are a paradigm shift in how we generate low-discrepancy point sets. Unlike traditional methods, where points are generated independently, GNNs allow points to 'chat' with one another so the network learns to place points in a way that reduces clustering and gaps — common issues with typical approaches.”

Going forward, the team plans to make MPMC points even more accessible to everyone, addressing the current limitation of training a new GNN for every fixed number of points and dimensions.

“Much of applied mathematics uses continuously varying quantities, but computation typically allows us to only use a finite number of points,” says Art B. Owen, Stanford University professor of statistics, who wasn’t involved in the research. “The century-plus-old field of discrepancy uses abstract algebra and number theory to define effective sampling points. This paper uses graph neural networks to find input points with low discrepancy compared to a continuous distribution. That approach already comes very close to the best-known low-discrepancy point sets in small problems and is showing great promise for a 32-dimensional integral from computational finance. We can expect this to be the first of many efforts to use neural methods to find good input points for numerical computation.”

Rusch and Rus wrote the paper with University of Waterloo researcher Nathan Kirk, Oxford University’s DeepMind Professor of AI and former CSAIL affiliate Michael Bronstein, and University of Waterloo Statistics and Actuarial Science Professor Christiane Lemieux. Their research was supported, in part, by the AI2050 program at Schmidt Sciences, Boeing, the United States Air Force Research Laboratory and the United States Air Force Artificial Intelligence Accelerator, the Swiss National Science Foundation, Natural Science and Engineering Research Council of Canada, and an EPSRC Turing AI World-Leading Research Fellowship. 

Have you ever wanted to travel through time to see what your future self might be like? Now, thanks to the power of generative AI, you can.

Researchers from MIT and elsewhere created a system that enables users to have an online, text-based conversation with an AI-generated simulation of their potential future self.

Dubbed Future You, the system is aimed at helping young people improve their sense of future self-continuity, a psychological concept that describes how connected a person feels with their future self.

Research has shown that a stronger sense of future self-continuity can positively influence how people make long-term decisions, from one’s likelihood to contribute to financial savings to their focus on achieving academic success.

Future You utilizes a large language model that draws on information provided by the user to generate a relatable, virtual version of the individual at age 60. This simulated future self can answer questions about what someone’s life in the future could be like, as well as offer advice or insights on the path they could follow.

In an initial user study, the researchers found that after interacting with Future You for about half an hour, people reported decreased anxiety and felt a stronger sense of connection with their future selves.

“We don’t have a real time machine yet, but AI can be a type of virtual time machine. We can use this simulation to help people think more about the consequences of the choices they are making today,” says Pat Pataranutaporn, a recent Media Lab doctoral graduate who is actively developing a program to advance human-AI interaction research at MIT, and co-lead author of a paper on Future You.

Pataranutaporn is joined on the paper by co-lead authors Kavin Winson, a researcher at KASIKORN Labs; and Peggy Yin, a Harvard University undergraduate; as well as Auttasak Lapapirojn and Pichayoot Ouppaphan of KASIKORN Labs; and senior authors Monchai Lertsutthiwong, head of AI research at the KASIKORN Business-Technology Group; Pattie Maes, the Germeshausen Professor of Media, Arts, and Sciences and head of the Fluid Interfaces group at MIT, and Hal Hershfield, professor of marketing, behavioral decision making, and psychology at the University of California at Los Angeles. The research will be presented at the IEEE Conference on Frontiers in Education.

A realistic simulation

Studies about conceptualizing one’s future self go back to at least the 1960s. One early method aimed at improving future self-continuity had people write letters to their future selves. More recently, researchers utilized virtual reality goggles to help people visualize future versions of themselves.

But none of these methods were very interactive, limiting the impact they could have on a user.

With the advent of generative AI and large language models like ChatGPT, the researchers saw an opportunity to make a simulated future self that could discuss someone’s actual goals and aspirations during a normal conversation.

“The system makes the simulation very realistic. Future You is much more detailed than what a person could come up with by just imagining their future selves,” says Maes.

Users begin by answering a series of questions about their current lives, things that are important to them, and goals for the future.

The AI system uses this information to create what the researchers call “future self memories” which provide a backstory the model pulls from when interacting with the user.

For instance, the chatbot could talk about the highlights of someone’s future career or answer questions about how the user overcame a particular challenge. This is possible because ChatGPT has been trained on extensive data involving people talking about their lives, careers, and good and bad experiences.

The user engages with the tool in two ways: through introspection, when they consider their life and goals as they construct their future selves, and retrospection, when they contemplate whether the simulation reflects who they see themselves becoming, says Yin.

“You can imagine Future You as a story search space. You have a chance to hear how some of your experiences, which may still be emotionally charged for you now, could be metabolized over the course of time,” she says.

To help people visualize their future selves, the system generates an age-progressed photo of the user. The chatbot is also designed to provide vivid answers using phrases like “when I was your age,” so the simulation feels more like an actual future version of the individual.

The ability to take advice from an older version of oneself, rather than a generic AI, can have a stronger positive impact on a user contemplating an uncertain future, Hershfield says.

“The interactive, vivid components of the platform give the user an anchor point and take something that could result in anxious rumination and make it more concrete and productive,” he adds.

But that realism could backfire if the simulation moves in a negative direction. To prevent this, they ensure Future You cautions users that it shows only one potential version of their future self, and they have the agency to change their lives. Providing alternate answers to the questionnaire yields a totally different conversation.

“This is not a prophesy, but rather a possibility,” Pataranutaporn says.

Aiding self-development

To evaluate Future You, they conducted a user study with 344 individuals. Some users interacted with the system for 10-30 minutes, while others either interacted with a generic chatbot or only filled out surveys.

Participants who used Future You were able to build a closer relationship with their ideal future selves, based on a statistical analysis of their responses. These users also reported less anxiety about the future after their interactions. In addition, Future You users said the conversation felt sincere and that their values and beliefs seemed consistent in their simulated future identities.

“This work forges a new path by taking a well-established psychological technique to visualize times to come — an avatar of the future self — with cutting edge AI. This is exactly the type of work academics should be focusing on as technology to build virtual self models merges with large language models,” says Jeremy Bailenson, the Thomas More Storke Professor of Communication at Stanford University, who was not involved with this research.

Building off the results of this initial user study, the researchers continue to fine-tune the ways they establish context and prime users so they have conversations that help build a stronger sense of future self-continuity.

“We want to guide the user to talk about certain topics, rather than asking their future selves who the next president will be,” Pataranutaporn says.

They are also adding safeguards to prevent people from misusing the system. For instance, one could imagine a company creating a “future you” of a potential customer who achieves some great outcome in life because they purchased a particular product.

Moving forward, the researchers want to study specific applications of Future You, perhaps by enabling people to explore different careers or visualize how their everyday choices could impact climate change.

They are also gathering data from the Future You pilot to better understand how people use the system.

“We don’t want people to become dependent on this tool. Rather, we hope it is a meaningful experience that helps them see themselves and the world differently, and helps with self-development,” Maes says.

The researchers acknowledge the support of Thanawit Prasongpongchai, a designer at KBTG and visiting scientist at the Media Lab.

The MIT Center for Transportation and Logistics (MIT CTL) and the Council of Supply Chain Management Professionals (CSCMP) have released the 2024 State of Supply Chain Sustainability report, marking the fifth edition of this influential research. The report highlights how supply chain sustainability practices have evolved over the past five years, assessing their global implementation and implications for industries, professionals, and the environment.

This year’s report is based on four years of comprehensive international surveys with responses from over 7,000 supply chain professionals representing more than 80 countries, coupled with insights from executive interviews. It explores how external pressures on firms, such as the growing investor demand and climate regulations, are driving sustainability initiatives. However, it also reveals persistent gaps between companies’ sustainability goals and the actual investments required to achieve them.

"Over the past five years, we have seen supply chains face unprecedented global challenges. While companies have made strides, our analysis shows that many are still struggling to align their sustainability ambitions with real progress, particularly when it comes to tackling Scope 3 emissions," says Josué Velázquez Martínez, MIT CTL research scientist and lead investigator. "Scope 3 emissions, which account for the vast majority of a company’s carbon footprint, remain a major hurdle due to the complexity of tracking emissions from indirect supply chain activities. The margin of error of the most common approach to estimate emissions are drastic, which disincentivizes companies to make more sustainable choices at the expense of investing in green alternatives."

Among the key findings:

• Increased pressure from investors: Over five years, pressure from investors to improve supply chain sustainability has grown by 25 percent, making it the fastest-growing driver of sustainability efforts.
• Lack of readiness for net-zero goals: Although 67 percent of firms surveyed do not have a net-zero goal in place, those that do are often unprepared to meet them, especially when it comes to measuring and reducing Scope 3 emissions.
• Company response to sustainability efforts in times of crisis: Companies react to different types of crises differently in regards to staying on track with their sustainable goals, whether it is a network disruption like the Covid-19 pandemic or economic turbulence.
• Challenges with Scope 3 emissions: Despite significant efforts, Scope 3 emissions — which can account for up to 75 percent of a company’s total emissions — continue to be the most difficult to track and manage, due to the complexity of supplier networks and inconsistent data-sharing practices.

Mark Baxa, president and CEO of CSCMP, emphasized the importance of collaboration: "Businesses and consumers alike are putting pressure on us to source and supply products to live up to their social and environmental standards. The State of Supply Chain Sustainability 2024 provides a thorough analysis of our current understanding, along with valuable insights on how to improve our Scope 3 emissions accounting to have a greater impact on lowering our emissions."

The report also underscores the importance of technological innovations, such as machine learning, advanced data analytics, and standardization to improve the accuracy of emissions tracking and help firms make data-driven sustainability decisions.

The 2024 State of Supply Chain Sustainability can be accessed online or in PDF format at sustainable.mit.edu.

The MIT CTL is a world leader in supply chain management research and education, with over 50 years of expertise. The center's work spans industry partnerships, cutting-edge research, and the advancement of sustainable supply chain practices. CSCMP is the leading global association for supply chain professionals. Established in 1963, CSCMP provides its members with education, research, and networking opportunities to advance the field of supply chain management.

In 1994, Florida jewelry designer Diana Duyser discovered what she believed to be the Virgin Mary’s image in a grilled cheese sandwich, which she preserved and later auctioned for $28,000. But how much do we really understand about pareidolia, the phenomenon of seeing faces and patterns in objects when they aren’t really there? 

A new study from the MIT Computer Science and Artificial Intelligence Laboratory (CSAIL) delves into this phenomenon, introducing an extensive, human-labeled dataset of 5,000 pareidolic images, far surpassing previous collections. Using this dataset, the team discovered several surprising results about the differences between human and machine perception, and how the ability to see faces in a slice of toast might have saved your distant relatives’ lives.

“Face pareidolia has long fascinated psychologists, but it’s been largely unexplored in the computer vision community,” says Mark Hamilton, MIT PhD student in electrical engineering and computer science, CSAIL affiliate, and lead researcher on the work. “We wanted to create a resource that could help us understand how both humans and AI systems process these illusory faces.”

So what did all of these fake faces reveal? For one, AI models don’t seem to recognize pareidolic faces like we do. Surprisingly, the team found that it wasn’t until they trained algorithms to recognize animal faces that they became significantly better at detecting pareidolic faces. This unexpected connection hints at a possible evolutionary link between our ability to spot animal faces — crucial for survival — and our tendency to see faces in inanimate objects. “A result like this seems to suggest that pareidolia might not arise from human social behavior, but from something deeper: like quickly spotting a lurking tiger, or identifying which way a deer is looking so our primordial ancestors could hunt,” says Hamilton.

Another intriguing discovery is what the researchers call the “Goldilocks Zone of Pareidolia,” a class of images where pareidolia is most likely to occur. “There’s a specific range of visual complexity where both humans and machines are most likely to perceive faces in non-face objects,” William T. Freeman, MIT professor of electrical engineering and computer science and principal investigator of the project says. “Too simple, and there’s not enough detail to form a face. Too complex, and it becomes visual noise.”

To uncover this, the team developed an equation that models how people and algorithms detect illusory faces.  When analyzing this equation, they found a clear “pareidolic peak” where the likelihood of seeing faces is highest, corresponding to images that have “just the right amount” of complexity. This predicted “Goldilocks zone” was then validated in tests with both real human subjects and AI face detection systems.

This new dataset, “Faces in Things,” dwarfs those of previous studies that typically used only 20-30 stimuli. This scale allowed the researchers to explore how state-of-the-art face detection algorithms behaved after fine-tuning on pareidolic faces, showing that not only could these algorithms be edited to detect these faces, but that they could also act as a silicon stand-in for our own brain, allowing the team to ask and answer questions about the origins of pareidolic face detection that are impossible to ask in humans. 

To build this dataset, the team curated approximately 20,000 candidate images from the LAION-5B dataset, which were then meticulously labeled and judged by human annotators. This process involved drawing bounding boxes around perceived faces and answering detailed questions about each face, such as the perceived emotion, age, and whether the face was accidental or intentional. “Gathering and annotating thousands of images was a monumental task,” says Hamilton. “Much of the dataset owes its existence to my mom,” a retired banker, “who spent countless hours lovingly labeling images for our analysis.”

The study also has potential applications in improving face detection systems by reducing false positives, which could have implications for fields like self-driving cars, human-computer interaction, and robotics. The dataset and models could also help areas like product design, where understanding and controlling pareidolia could create better products. “Imagine being able to automatically tweak the design of a car or a child’s toy so it looks friendlier, or ensuring a medical device doesn’t inadvertently appear threatening,” says Hamilton.

“It’s fascinating how humans instinctively interpret inanimate objects with human-like traits. For instance, when you glance at an electrical socket, you might immediately envision it singing, and you can even imagine how it would ‘move its lips.’ Algorithms, however, don’t naturally recognize these cartoonish faces in the same way we do,” says Hamilton. “This raises intriguing questions: What accounts for this difference between human perception and algorithmic interpretation? Is pareidolia beneficial or detrimental? Why don’t algorithms experience this effect as we do? These questions sparked our investigation, as this classic psychological phenomenon in humans had not been thoroughly explored in algorithms.”

As the researchers prepare to share their dataset with the scientific community, they’re already looking ahead. Future work may involve training vision-language models to understand and describe pareidolic faces, potentially leading to AI systems that can engage with visual stimuli in more human-like ways.

“This is a delightful paper! It is fun to read and it makes me think. Hamilton et al. propose a tantalizing question: Why do we see faces in things?” says Pietro Perona, the Allen E. Puckett Professor of Electrical Engineering at Caltech, who was not involved in the work. “As they point out, learning from examples, including animal faces, goes only half-way to explaining the phenomenon. I bet that thinking about this question will teach us something important about how our visual system generalizes beyond the training it receives through life.”

Hamilton and Freeman’s co-authors include Simon Stent, staff research scientist at the Toyota Research Institute; Ruth Rosenholtz, principal research scientist in the Department of Brain and Cognitive Sciences, NVIDIA research scientist, and former CSAIL member; and CSAIL affiliates postdoc Vasha DuTell, Anne Harrington MEng ’23, and Research Scientist Jennifer Corbett. Their work was supported, in part, by the National Science Foundation and the CSAIL MEnTorEd Opportunities in Research (METEOR) Fellowship, while being sponsored by the United States Air Force Research Laboratory and the United States Air Force Artificial Intelligence Accelerator. The MIT SuperCloud and Lincoln Laboratory Supercomputing Center provided HPC resources for the researchers’ results.

This work is being presented this week at the European Conference on Computer Vision.

Imagine having to straighten up a messy kitchen, starting with a counter littered with sauce packets. If your goal is to wipe the counter clean, you might sweep up the packets as a group. If, however, you wanted to first pick out the mustard packets before throwing the rest away, you would sort more discriminately, by sauce type. And if, among the mustards, you had a hankering for Grey Poupon, finding this specific brand would entail a more careful search.

MIT engineers have developed a method that enables robots to make similarly intuitive, task-relevant decisions.

The team’s new approach, named Clio, enables a robot to identify the parts of a scene that matter, given the tasks at hand. With Clio, a robot takes in a list of tasks described in natural language and, based on those tasks, it then determines the level of granularity required to interpret its surroundings and “remember” only the parts of a scene that are relevant.

In real experiments ranging from a cluttered cubicle to a five-story building on MIT’s campus, the team used Clio to automatically segment a scene at different levels of granularity, based on a set of tasks specified in natural-language prompts such as “move rack of magazines” and “get first aid kit.”

The team also ran Clio in real-time on a quadruped robot. As the robot explored an office building, Clio identified and mapped only those parts of the scene that related to the robot’s tasks (such as retrieving a dog toy while ignoring piles of office supplies), allowing the robot to grasp the objects of interest.

Clio is named after the Greek muse of history, for its ability to identify and remember only the elements that matter for a given task. The researchers envision that Clio would be useful in many situations and environments in which a robot would have to quickly survey and make sense of its surroundings in the context of its given task.

“Search and rescue is the motivating application for this work, but Clio can also power domestic robots and robots working on a factory floor alongside humans,” says Luca Carlone, associate professor in MIT’s Department of Aeronautics and Astronautics (AeroAstro), principal investigator in the Laboratory for Information and Decision Systems (LIDS), and director of the MIT SPARK Laboratory. “It’s really about helping the robot understand the environment and what it has to remember in order to carry out its mission.”

The team details their results in a study appearing today in the journal Robotics and Automation Letters. Carlone’s co-authors include members of the SPARK Lab: Dominic Maggio, Yun Chang, Nathan Hughes, and Lukas Schmid; and members of MIT Lincoln Laboratory: Matthew Trang, Dan Griffith, Carlyn Dougherty, and Eric Cristofalo.

Open fields

Huge advances in the fields of computer vision and natural language processing have enabled robots to identify objects in their surroundings. But until recently, robots were only able to do so in “closed-set” scenarios, where they are programmed to work in a carefully curated and controlled environment, with a finite number of objects that the robot has been pretrained to recognize.

In recent years, researchers have taken a more “open” approach to enable robots to recognize objects in more realistic settings. In the field of open-set recognition, researchers have leveraged deep-learning tools to build neural networks that can process billions of images from the internet, along with each image’s associated text (such as a friend’s Facebook picture of a dog, captioned “Meet my new puppy!”).

From millions of image-text pairs, a neural network learns from, then identifies, those segments in a scene that are characteristic of certain terms, such as a dog. A robot can then apply that neural network to spot a dog in a totally new scene.

But a challenge still remains as to how to parse a scene in a useful way that is relevant for a particular task.

“Typical methods will pick some arbitrary, fixed level of granularity for determining how to fuse segments of a scene into what you can consider as one ‘object,’” Maggio says. “However, the granularity of what you call an ‘object’ is actually related to what the robot has to do. If that granularity is fixed without considering the tasks, then the robot may end up with a map that isn’t useful for its tasks.”

Information bottleneck

With Clio, the MIT team aimed to enable robots to interpret their surroundings with a level of granularity that can be automatically tuned to the tasks at hand.

For instance, given a task of moving a stack of books to a shelf, the robot should be able to  determine that the entire stack of books is the task-relevant object. Likewise, if the task were to move only the green book from the rest of the stack, the robot should distinguish the green book as a single target object and disregard the rest of the scene — including the other books in the stack.

The team’s approach combines state-of-the-art computer vision and large language models comprising neural networks that make connections among millions of open-source images and semantic text. They also incorporate mapping tools that automatically split an image into many small segments, which can be fed into the neural network to determine if certain segments are semantically similar. The researchers then leverage an idea from classic information theory called the “information bottleneck,” which they use to compress a number of image segments in a way that picks out and stores segments that are semantically most relevant to a given task.

“For example, say there is a pile of books in the scene and my task is just to get the green book. In that case we push all this information about the scene through this bottleneck and end up with a cluster of segments that represent the green book,” Maggio explains. “All the other segments that are not relevant just get grouped in a cluster which we can simply remove. And we’re left with an object at the right granularity that is needed to support my task.”

The researchers demonstrated Clio in different real-world environments.

“What we thought would be a really no-nonsense experiment would be to run Clio in my apartment, where I didn’t do any cleaning beforehand,” Maggio says.

The team drew up a list of natural-language tasks, such as “move pile of clothes” and then applied Clio to images of Maggio’s cluttered apartment. In these cases, Clio was able to quickly segment scenes of the apartment and feed the segments through the Information Bottleneck algorithm to identify those segments that made up the pile of clothes.

They also ran Clio on Boston Dynamic’s quadruped robot, Spot. They gave the robot a list of tasks to complete, and as the robot explored and mapped the inside of an office building, Clio ran in real-time on an on-board computer mounted to Spot, to pick out segments in the mapped scenes that visually relate to the given task. The method generated an overlaying map showing just the target objects, which the robot then used to approach the identified objects and physically complete the task.

“Running Clio in real-time was a big accomplishment for the team,” Maggio says. “A lot of prior work can take several hours to run.”

Going forward, the team plans to adapt Clio to be able to handle higher-level tasks and build upon recent advances in photorealistic visual scene representations.

“We’re still giving Clio tasks that are somewhat specific, like ‘find deck of cards,’” Maggio says. “For search and rescue, you need to give it more high-level tasks, like ‘find survivors,’ or ‘get power back on.’ So, we want to get to a more human-level understanding of how to accomplish more complex tasks.”

This research was supported, in part, by the U.S. National Science Foundation, the Swiss National Science Foundation, MIT Lincoln Laboratory, the U.S. Office of Naval Research, and the U.S. Army Research Lab Distributed and Collaborative Intelligent Systems and Technology Collaborative Research Alliance.

Deep-learning models are being used in many fields, from health care diagnostics to financial forecasting. However, these models are so computationally intensive that they require the use of powerful cloud-based servers.

This reliance on cloud computing poses significant security risks, particularly in areas like health care, where hospitals may be hesitant to use AI tools to analyze confidential patient data due to privacy concerns.

To tackle this pressing issue, MIT researchers have developed a security protocol that leverages the quantum properties of light to guarantee that data sent to and from a cloud server remain secure during deep-learning computations.

By encoding data into the laser light used in fiber optic communications systems, the protocol exploits the fundamental principles of quantum mechanics, making it impossible for attackers to copy or intercept the information without detection.

Moreover, the technique guarantees security without compromising the accuracy of the deep-learning models. In tests, the researcher demonstrated that their protocol could maintain 96 percent accuracy while ensuring robust security measures.

“Deep learning models like GPT-4 have unprecedented capabilities but require massive computational resources. Our protocol enables users to harness these powerful models without compromising the privacy of their data or the proprietary nature of the models themselves,” says Kfir Sulimany, an MIT postdoc in the Research Laboratory for Electronics (RLE) and lead author of a paper on this security protocol.

Sulimany is joined on the paper by Sri Krishna Vadlamani, an MIT postdoc; Ryan Hamerly, a former postdoc now at NTT Research, Inc.; Prahlad Iyengar, an electrical engineering and computer science (EECS) graduate student; and senior author Dirk Englund, a professor in EECS, principal investigator of the Quantum Photonics and Artificial Intelligence Group and of RLE. The research was recently presented at Annual Conference on Quantum Cryptography.

A two-way street for security in deep learning

The cloud-based computation scenario the researchers focused on involves two parties — a client that has confidential data, like medical images, and a central server that controls a deep learning model.

The client wants to use the deep-learning model to make a prediction, such as whether a patient has cancer based on medical images, without revealing information about the patient.

In this scenario, sensitive data must be sent to generate a prediction. However, during the process the patient data must remain secure.

Also, the server does not want to reveal any parts of the proprietary model that a company like OpenAI spent years and millions of dollars building.

“Both parties have something they want to hide,” adds Vadlamani.

In digital computation, a bad actor could easily copy the data sent from the server or the client.

Quantum information, on the other hand, cannot be perfectly copied. The researchers leverage this property, known as the no-cloning principle, in their security protocol.

For the researchers’ protocol, the server encodes the weights of a deep neural network into an optical field using laser light.

A neural network is a deep-learning model that consists of layers of interconnected nodes, or neurons, that perform computation on data. The weights are the components of the model that do the mathematical operations on each input, one layer at a time. The output of one layer is fed into the next layer until the final layer generates a prediction.

The server transmits the network’s weights to the client, which implements operations to get a result based on their private data. The data remain shielded from the server.

At the same time, the security protocol allows the client to measure only one result, and it prevents the client from copying the weights because of the quantum nature of light.

Once the client feeds the first result into the next layer, the protocol is designed to cancel out the first layer so the client can’t learn anything else about the model.

“Instead of measuring all the incoming light from the server, the client only measures the light that is necessary to run the deep neural network and feed the result into the next layer. Then the client sends the residual light back to the server for security checks,” Sulimany explains.

Due to the no-cloning theorem, the client unavoidably applies tiny errors to the model while measuring its result. When the server receives the residual light from the client, the server can measure these errors to determine if any information was leaked. Importantly, this residual light is proven to not reveal the client data.

A practical protocol

Modern telecommunications equipment typically relies on optical fibers to transfer information because of the need to support massive bandwidth over long distances. Because this equipment already incorporates optical lasers, the researchers can encode data into light for their security protocol without any special hardware.

When they tested their approach, the researchers found that it could guarantee security for server and client while enabling the deep neural network to achieve 96 percent accuracy.

The tiny bit of information about the model that leaks when the client performs operations amounts to less than 10 percent of what an adversary would need to recover any hidden information. Working in the other direction, a malicious server could only obtain about 1 percent of the information it would need to steal the client’s data.

“You can be guaranteed that it is secure in both ways — from the client to the server and from the server to the client,” Sulimany says.

“A few years ago, when we developed our demonstration of distributed machine learning inference between MIT’s main campus and MIT Lincoln Laboratory, it dawned on me that we could do something entirely new to provide physical-layer security, building on years of quantum cryptography work that had also been shown on that testbed,” says Englund. “However, there were many deep theoretical challenges that had to be overcome to see if this prospect of privacy-guaranteed distributed machine learning could be realized. This didn’t become possible until Kfir joined our team, as Kfir uniquely understood the experimental as well as theory components to develop the unified framework underpinning this work.”

In the future, the researchers want to study how this protocol could be applied to a technique called federated learning, where multiple parties use their data to train a central deep-learning model. It could also be used in quantum operations, rather than the classical operations they studied for this work, which could provide advantages in both accuracy and security.

“This work combines in a clever and intriguing way techniques drawing from fields that do not usually meet, in particular, deep learning and quantum key distribution. By using methods from the latter, it adds a security layer to the former, while also allowing for what appears to be a realistic implementation. This can be interesting for preserving privacy in distributed architectures. I am looking forward to seeing how the protocol behaves under experimental imperfections and its practical realization,” says Eleni Diamanti, a CNRS research director at Sorbonne University in Paris, who was not involved with this work.

This work was supported, in part, by the Israeli Council for Higher Education and the Zuckerman STEM Leadership Program.

Mars wasn’t always the cold desert we see today. There’s increasing evidence that water once flowed on the Red Planet’s surface, billions of years ago. And if there was water, there must also have been a thick atmosphere to keep that water from freezing. But sometime around 3.5 billion years ago, the water dried up, and the air, once heavy with carbon dioxide, dramatically thinned, leaving only the wisp of an atmosphere that clings to the planet today.

Where exactly did Mars’ atmosphere go? This question has been a central mystery of Mars’ 4.6-billion-year history.

For two MIT geologists, the answer may lie in the planet’s clay. In a paper appearing today in Science Advances, they propose that much of Mars’ missing atmosphere could be locked up in the planet’s clay-covered crust.

The team makes the case that, while water was present on Mars, the liquid could have trickled through certain rock types and set off a slow chain of reactions that progressively drew carbon dioxide out of the atmosphere and converted it into methane — a form of carbon that could be stored for eons in the planet’s clay surface.

Similar processes occur in some regions on Earth. The researchers used their knowledge of interactions between rocks and gases on Earth and applied that to how similar processes could play out on Mars. They found that, given how much clay is estimated to cover Mars’ surface, the planet’s clay could hold up to 1.7 bar of carbon dioxide, which would be equivalent to around 80 percent of the planet’s initial, early atmosphere.

It’s possible that this sequestered Martian carbon could one day be recovered and converted into propellant to fuel future missions between Mars and Earth, the researchers propose.

“Based on our findings on Earth, we show that similar processes likely operated on Mars, and that copious amounts of atmospheric CO₂ could have transformed to methane and been sequestered in clays,” says study author Oliver Jagoutz, professor of geology in MIT’s Department of Earth, Atmospheric and Planetary Sciences (EAPS). “This methane could still be present and maybe even used as an energy source on Mars in the future.”

The study’s lead author is recent EAPS graduate Joshua Murray PhD ’24.

In the folds

Jagoutz’ group at MIT seeks to identify the geologic processes and interactions that drive the evolution of Earth’s lithosphere — the hard and brittle outer layer that includes the crust and upper mantle, where tectonic plates lie.

In 2023, he and Murray focused on a type of surface clay mineral called smectite, which is known to be a highly effective trap for carbon. Within a single grain of smectite are a multitude of folds, within which carbon can sit undisturbed for billions of years. They showed that smectite on Earth was likely a product of tectonic activity, and that, once exposed at the surface, the clay minerals acted to draw down and store enough carbon dioxide from the atmosphere to cool the planet over millions of years.

Soon after the team reported their results, Jagoutz happened to look at a map of the surface of Mars and realized that much of that planet’s surface was covered in the same smectite clays. Could the clays have had a similar carbon-trapping effect on Mars, and if so, how much carbon could the clays hold?

“We know this process happens, and it is well-documented on Earth. And these rocks and clays exist on Mars,” Jagoutz says. “So, we wanted to try and connect the dots.”

“Every nook and cranny”

Unlike on Earth, where smectite is a consequence of continental plates shifting and uplifting to bring rocks from the mantle to the surface, there is no such tectonic activity on Mars. The team looked for ways in which the clays could have formed on Mars, based on what scientists know of the planet’s history and composition.

For instance, some remote measurements of Mars’ surface suggest that at least part of the planet’s crust contains ultramafic igneous rocks, similar to those that produce smectites through weathering on Earth. Other observations reveal geologic patterns similar to terrestrial rivers and tributaries, where water could have flowed and reacted with the underlying rock.

Jagoutz and Murray wondered whether water could have reacted with Mars’ deep ultramafic rocks in a way that would produce the clays that cover the surface today. They developed a simple model of rock chemistry, based on what is known of how igneous rocks interact with their environment on Earth.

They applied this model to Mars, where scientists believe the crust is mostly made up of igneous rock that is rich in the mineral olivine. The team used the model to estimate the changes that olivine-rich rock might undergo, assuming that water existed on the surface for at least a billion years, and the atmosphere was thick with carbon dioxide.

“At this time in Mars’ history, we think CO₂ is everywhere, in every nook and cranny, and water percolating through the rocks is full of CO₂ too,” Murray says.

Over about a billion years, water trickling through the crust would have slowly reacted with olivine — a mineral that is rich in a reduced form of iron. Oxygen molecules in water would have bound to the iron, releasing hydrogen as a result and forming the red oxidized iron which gives the planet its iconic color. This free hydrogen would then have combined with carbon dioxide in the water, to form methane. As this reaction progressed over time, olivine would have slowly transformed into another type of iron-rich rock known as serpentine, which then continued to react with water to form smectite.

“These smectite clays have so much capacity to store carbon,” Murray says. “So then we used existing knowledge of how these minerals are stored in clays on Earth, and extrapolate to say, if the Martian surface has this much clay in it, how much methane can you store in those clays?”

He and Jagoutz found that if Mars is covered in a layer of smectite that is 1,100 meters deep, this amount of clay could store a huge amount of methane, equivalent to most of the carbon dioxide in the atmosphere that is thought to have disappeared since the planet dried up.

“We find that estimates of global clay volumes on Mars are consistent with a significant fraction of Mars’ initial CO₂ being sequestered as organic compounds within the clay-rich crust,” Murray says. “In some ways, Mars’ missing atmosphere could be hiding in plain sight.”

“Where the CO₂ went from an early, thicker atmosphere is a fundamental question in the history of the Mars atmosphere, its climate, and the habitability by microbes,” says Bruce Jakosky, professor emeritus of geology at the University of Colorado and principal investigator on the Mars Atmosphere and Volatile Evolution (MAVEN) mission, which has been orbiting and studying Mars’ upper atmosphere since 2014. Jakosky was not involved with the current study. “Murray and Jagoutz examine the chemical interaction of rocks with the atmosphere as a means of removing CO2. At the high end of our estimates of how much weathering has occurred, this could be a major process in removing CO₂ from Mars’ early atmosphere.”

This work was supported, in part, by the National Science Foundation.

When summer temperatures spike, so does our vulnerability to heat-related illness or even death. For the most part, people can take measures to reduce their heat exposure by opening a window, turning up the air conditioning, or simply getting a glass of water. But for people who are incarcerated, freedom to take such measures is often not an option. Prison populations therefore are especially vulnerable to heat exposure, due to their conditions of confinement.

A new study by MIT researchers examines summertime heat exposure in prisons across the United States and identifies characteristics within prison facilities that can further contribute to a population’s vulnerability to summer heat.

The study’s authors used high-spatial-resolution air temperature data to determine the daily average outdoor temperature for each of 1,614 prisons in the U.S., for every summer between the years 1990 and 2023. They found that the prisons that are exposed to the most extreme heat are located in the southwestern U.S., while prisons with the biggest changes in summertime heat, compared to the historical record, are in the Pacific Northwest, the Northeast, and parts of the Midwest.

Those findings are not entirely unique to prisons, as any non-prison facility or community in the same geographic locations would be exposed to similar outdoor air temperatures. But the team also looked at characteristics specific to prison facilities that could further exacerbate an incarcerated person’s vulnerability to heat exposure. They identified nine such facility-level characteristics, such as highly restricted movement, poor staffing, and inadequate mental health treatment. People living and working in prisons with any one of these characteristics may experience compounded risk to summertime heat. 

The team also looked at the demographics of 1,260 prisons in their study and found that the prisons with higher heat exposure on average also had higher proportions of non-white and Hispanic populations. The study, appearing today in the journal GeoHealth, provides policymakers and community leaders with ways to estimate, and take steps to address, a prison population’s heat risk, which they anticipate could worsen with climate change.

“This isn’t a problem because of climate change. It’s becoming a worse problem because of climate change,” says study lead author Ufuoma Ovienmhada SM ’20, PhD ’24, a graduate of the MIT Media Lab, who recently completed her doctorate in MIT’s Department of Aeronautics and Astronautics (AeroAstro). “A lot of these prisons were not built to be comfortable or humane in the first place. Climate change is just aggravating the fact that prisons are not designed to enable incarcerated populations to moderate their own exposure to environmental risk factors such as extreme heat.”

The study’s co-authors include Danielle Wood ’04, SM ’08, PhD ’12, MIT associate professor of media arts and sciences, and of AeroAstro; and Brent Minchew, MIT associate professor of geophysics in the Department of Earth, Atmospheric and Planetary Sciences; along with Ahmed Diongue ’24, Mia Hines-Shanks of Grinnell College, and Michael Krisch of Columbia University.

Environmental intersections

The new study is an extension of work carried out at the Media Lab, where Wood leads the Space Enabled research group. The group aims to advance social and environmental justice issues through the use of satellite data and other space-enabled technologies.

The group’s motivation to look at heat exposure in prisons came in 2020 when, as co-president of MIT’s Black Graduate Student Union, Ovienmhada took part in community organizing efforts following the murder of George Floyd by Minneapolis police.

“We started to do more organizing on campus around policing and reimagining public safety. Through that lens I learned more about police and prisons as interconnected systems, and came across this intersection between prisons and environmental hazards,” says Ovienmhada, who is leading an effort to map the various environmental hazards that prisons, jails, and detention centers face. “In terms of environmental hazards, extreme heat causes some of the most acute impacts for incarcerated people.”

She, Wood, and their colleagues set out to use Earth observation data to characterize U.S. prison populations’ vulnerability, or their risk of experiencing negative impacts, from heat.

The team first looked through a database maintained by the U.S. Department of Homeland Security that lists the location and boundaries of carceral facilities in the U.S. From the database’s more than 6,000 prisons, jails, and detention centers, the researchers highlighted 1,614 prison-specific facilities, which together incarcerate nearly 1.4 million people, and employ about 337,000 staff.

They then looked to Daymet, a detailed weather and climate database that tracks daily temperatures across the United States, at a 1-kilometer resolution. For each of the 1,614 prison locations, they mapped the daily outdoor temperature, for every summer between the years 1990 to 2023, noting that the majority of current state and federal correctional facilities in the U.S. were built by 1990.

The team also obtained U.S. Census data on each facility’s demographic and facility-level characteristics, such as prison labor activities and conditions of confinement. One limitation of the study that the researchers acknowledge is a lack of information regarding a prison’s climate control.

“There’s no comprehensive public resource where you can look up whether a facility has air conditioning,” Ovienmhada notes. “Even in facilities with air conditioning, incarcerated people may not have regular access to those cooling systems, so our measurements of outdoor air temperature may not be far off from reality.”

Heat factors

From their analysis, the researchers found that more than 98 percent of all prisons in the U.S. experienced at least 10 days in the summer that were hotter than every previous summer, on average, for a given location. Their analysis also revealed the most heat-exposed prisons, and the prisons that experienced the highest temperatures on average, were mostly in the Southwestern U.S. The researchers note that with the exception of New Mexico, the Southwest is a region where there are no universal air conditioning regulations in state-operated prisons.

“States run their own prison systems, and there is no uniformity of data collection or policy regarding air conditioning,” says Wood, who notes that there is some information on cooling systems in some states and individual prison facilities, but the data is sparse overall, and too inconsistent to include in the group’s nationwide study.

While the researchers could not incorporate air conditioning data, they did consider other facility-level factors that could worsen the effects that outdoor heat triggers. They looked through the scientific literature on heat, health impacts, and prison conditions, and focused on 17 measurable facility-level variables that contribute to heat-related health problems. These include factors such as overcrowding and understaffing.

“We know that whenever you’re in a room that has a lot of people, it’s going to feel hotter, even if there’s air conditioning in that environment,” Ovienmhada says. “Also, staffing is a huge factor. Facilities that don’t have air conditioning but still try to do heat risk-mitigation procedures might rely on staff to distribute ice or water every few hours. If that facility is understaffed or has neglectful staff, that may increase people’s susceptibility to hot days.”

The study found that prisons with any of nine of the 17 variables showed statistically significant greater heat exposures than the prisons without those variables. Additionally, if a prison exhibits any one of the nine variables, this could worsen people’s heat risk through the combination of elevated heat exposure and vulnerability. The variables, they say, could help state regulators and activists identify prisons to prioritize for heat interventions.

“The prison population is aging, and even if you’re not in a ‘hot state,’ every state has responsibility to respond,” Wood emphasizes. “For instance, areas in the Northwest, where you might expect to be temperate overall, have experienced a number of days in recent years of increasing heat risk. A few days out of the year can still be dangerous, particularly for a population with reduced agency to regulate their own exposure to heat.”

This work was supported, in part, by NASA, the MIT Media Lab, and MIT’s Institute for Data, Systems and Society’s Research Initiative on Combatting Systemic Racism.

The degree to which a surgical patient’s subconscious processing of pain, or “nociception,” is properly managed by their anesthesiologist will directly affect the degree of post-operative drug side effects they’ll experience and the need for further pain management they’ll require. But pain is a subjective feeling to measure, even when patients are awake, much less when they are unconscious. 

In a new study appearing in the Proceedings of the National Academy of Sciences, MIT and Massachusetts General Hospital (MGH) researchers describe a set of statistical models that objectively quantified nociception during surgery. Ultimately, they hope to help anesthesiologists optimize drug dose and minimize post-operative pain and side effects.

The new models integrate data meticulously logged over 18,582 minutes of 101 abdominal surgeries in men and women at MGH. Led by Sandya Subramanian PhD ’21, an assistant professor at the University of California at Berkeley and the University of California at San Francisco, the researchers collected and analyzed data from five physiological sensors as patients experienced a total of 49,878 distinct “nociceptive stimuli” (such as incisions or cautery). Moreover, the team recorded what drugs were administered, and how much and when, to factor in their effects on nociception or cardiovascular measures. They then used all the data to develop a set of statistical models that performed well in retrospectively indicating the body’s response to nociceptive stimuli.

The team’s goal is to furnish such accurate, objective, and physiologically principled information in real time to anesthesiologists who currently have to rely heavily on intuition and past experience in deciding how to administer pain-control drugs during surgery. If anesthesiologists give too much, patients can experience side effects ranging from nausea to delirium. If they give too little, patients may feel excessive pain after they awaken.

“Sandya’s work has helped us establish a principled way to understand and measure nociception (unconscious pain) during general anesthesia,” says study senior author Emery N. Brown, the Edward Hood Taplin Professor of Medical Engineering and Computational Neuroscience in The Picower Institute for Learning and Memory, the Institute for Medical Engineering and Science, and the Department of Brain and Cognitive Sciences at MIT. Brown is also an anesthesiologist at MGH and a professor at Harvard Medical School. “Our next objective is to make the insights that we have gained from Sandya’s studies reliable and practical for anesthesiologists to use during surgery.”

Surgery and statistics

The research began as Subramanian’s doctoral thesis project in Brown’s lab in 2017. The best prior attempts to objectively model nociception have either relied solely on the electrocardiogram (ECG, an indirect indicator of heart-rate variability) or other systems that may incorporate more than one measurement, but were either based on lab experiments using pain stimuli that do not compare in intensity to surgical pain or were validated by statistically aggregating just a few time points across multiple patients’ surgeries, Subramanian says.

“There’s no other place to study surgical pain except for the operating room,” Subramanian says. “We wanted to not only develop the algorithms using data from surgery, but also actually validate it in the context in which we want someone to use it. If we are asking them to track moment-to-moment nociception during an individual surgery, we need to validate it in that same way.”

So she and Brown worked to advance the state of the art by collecting multi-sensor data during the whole course of actual surgeries and by accounting for the confounding effects of the drugs administered. In that way, they hoped to develop a model that could make accurate predictions that remained valid for the same patient all the way through their operation.

Part of the improvements the team achieved arose from tracking patterns of heart rate and also skin conductance. Changes in both of these physiological factors can be indications of the body’s primal “fight or flight” response to nociception or pain, but some drugs used during surgery directly affect cardiovascular state, while skin conductance (or “EDA,” electrodermal activity) remains unaffected. The study measures not only ECG but also backs it up with PPG, an optical measure of heart rate (like the oxygen sensor on a smartwatch), because ECG signals can sometimes be made noisy by all the electrical equipment buzzing away in the operating room. Similarly, Subramanian backstopped EDA measures with measures of skin temperature to ensure that changes in skin conductance from sweat were because of nociception and not simply the patient being too warm. The study also tracked respiration.

Then the authors performed statistical analyses to develop physiologically relevant indices from each of the cardiovascular and skin conductance signals. And once each index was established, further statistical analysis enabled tracking the indices together to produce models that could make accurate, principled predictions of when nociception was occurring and the body’s response.

Nailing nociception

In four versions of the model, Subramanian “supervised” them by feeding them information on when actual nociceptive stimuli occurred so that they could then learn the association between the physiological measurements and the incidence of pain-inducing events. In some of these trained versions she left out drug information and in some versions she used different statistical approaches (either “linear regression” or “random forest”). In a fifth version of the model, based on a “state space” approach, she left it unsupervised, meaning it had to learn to infer moments of nociception purely from the physiological indices. She compared all five versions of her model to one of the current industry standards, an ECG-tracking model called ANI.

Each model’s output can be visualized as a graph plotting the predicted degree of nociception over time. ANI performs just above chance but is implemented in real-time. The unsupervised model performed better than ANI, though not quite as well as the supervised models. The best performing of those was one that incorporated drug information and used a “random forest” approach. Still, the authors note, the fact that the unsupervised model performed significantly better than chance suggests that there is indeed an objectively detectable signature of the body’s nociceptive state even when looking across different patients.

“A state space framework using multisensory physiological observations is effective in uncovering this implicit nociceptive state with a consistent definition across multiple subjects,” wrote Subramanian, Brown, and their co-authors. “This is an important step toward defining a metric to track nociception without including nociceptive ‘ground truth’ information, most practical for scalability and implementation in clinical settings.”

Indeed, the next steps for the research are to increase the data sampling and to further refine the models so that they can eventually be put into practice in the operating room. That will require enabling them to predict nociception in real time, rather than in post-hoc analysis. When that advance is made, that will enable anesthesiologists or intensivists to inform their pain drug dosing judgements. Further into the future, the model could inform closed-loop systems that automatically dose drugs under the anesthesiologist’s supervision.

“Our study is an important first step toward developing objective markers to track surgical nociception,” the authors concluded. “These markers will enable objective assessment of nociception in other complex clinical settings, such as the ICU [intensive care unit], as well as catalyze future development of closed-loop control systems for nociception.”

In addition to Subramanian and Brown, the paper’s other authors are Bryan Tseng, Marcela del Carmen, Annekathryn Goodman, Douglas Dahl, and Riccardo Barbieri.

Funding from The JPB Foundation; The Picower Institute; George J. Elbaum ’59, SM ’63, PhD ’67; Mimi Jensen; Diane B. Greene SM ’78; Mendel Rosenblum; Bill Swanson; Cathy and Lou Paglia; annual donors to the Anesthesia Initiative Fund; the National Science Foundation; and an MIT Office of Graduate Education Collabmore-Rogers Fellowship supported the research.

The pharmaceutical manufacturing industry has long struggled with the issue of monitoring the characteristics of a drying mixture, a critical step in producing medication and chemical compounds. At present, there are two noninvasive characterization approaches that are typically used: A sample is either imaged and individual particles are counted, or researchers use a scattered light to estimate the particle size distribution (PSD). The former is time-intensive and leads to increased waste, making the latter a more attractive option.

In recent years, MIT engineers and researchers developed a physics and machine learning-based scattered light approach that has been shown to improve manufacturing processes for pharmaceutical pills and powders, increasing efficiency and accuracy and resulting in fewer failed batches of products. A new open-access paper, “Non-invasive estimation of the powder size distribution from a single speckle image,” available in the journal Light: Science & Application, expands on this work, introducing an even faster approach. 

“Understanding the behavior of scattered light is one of the most important topics in optics,” says Qihang Zhang PhD ’23, an associate researcher at Tsinghua University. “By making progress in analyzing scattered light, we also invented a useful tool for the pharmaceutical industry. Locating the pain point and solving it by investigating the fundamental rule is the most exciting thing to the research team.”

The paper proposes a new PSD estimation method, based on pupil engineering, that reduces the number of frames needed for analysis. “Our learning-based model can estimate the powder size distribution from a single snapshot speckle image, consequently reducing the reconstruction time from 15 seconds to a mere 0.25 seconds,” the researchers explain.

“Our main contribution in this work is accelerating a particle size detection method by 60 times, with a collective optimization of both algorithm and hardware,” says Zhang. “This high-speed probe is capable to detect the size evolution in fast dynamical systems, providing a platform to study models of processes in pharmaceutical industry including drying, mixing and blending.”

The technique offers a low-cost, noninvasive particle size probe by collecting back-scattered light from powder surfaces. The compact and portable prototype is compatible with most of drying systems in the market, as long as there is an observation window. This online measurement approach may help control manufacturing processes, improving efficiency and product quality. Further, the previous lack of online monitoring prevented systematical study of dynamical models in manufacturing processes. This probe could bring a new platform to carry out series research and modeling for the particle size evolution.

This work, a successful collaboration between physicists and engineers, is generated from the MIT-Takeda program. Collaborators are affiliated with three MIT departments: Mechanical Engineering, Chemical Engineering, and Electrical Engineering and Computer Science. George Barbastathis, professor of mechanical engineering at MIT, is the article’s senior author.

One major reason why it has been difficult to develop an effective HIV vaccine is that the virus mutates very rapidly, allowing it to evade the antibody response generated by vaccines.

Several years ago, MIT researchers showed that administering a series of escalating doses of an HIV vaccine over a two-week period could help overcome a part of that challenge by generating larger quantities of neutralizing antibodies. However, a multidose vaccine regimen administered over a short time is not practical for mass vaccination campaigns.

In a new study, the researchers have now found that they can achieve a similar immune response with just two doses, given one week apart. The first dose, which is much smaller, prepares the immune system to respond more powerfully to the second, larger dose.

This study, which was performed by bringing together computational modeling and experiments in mice, used an HIV envelope protein as the vaccine. A single-dose version of this vaccine is now in clinical trials, and the researchers hope to establish another study group that will receive the vaccine on a two-dose schedule.

“By bringing together the physical and life sciences, we shed light on some basic immunological questions that helped develop this two-dose schedule to mimic the multiple-dose regimen,” says Arup Chakraborty, the John M. Deutch Institute Professor at MIT and a member of MIT’s Institute for Medical Engineering and Science and the Ragon Institute of MIT, MGH and Harvard University.

This approach may also generalize to vaccines for other diseases, Chakraborty notes.

Chakraborty and Darrell Irvine, a former MIT professor of biological engineering and materials science and engineering and member of the Koch Institute for Integrative Cancer Research, who is now a professor of immunology and microbiology at the Scripps Research Institute, are the senior authors of the study, which appears today in Science Immunology. The lead authors of the paper are Sachin Bhagchandani PhD ’23 and Leerang Yang PhD ’24.

Neutralizing antibodies

Each year, HIV infects more than 1 million people around the world, and some of those people do not have access to antiviral drugs. An effective vaccine could prevent many of those infections. One promising vaccine now in clinical trials consists of an HIV protein called an envelope trimer, along with a nanoparticle called SMNP. The nanoparticle, developed by Irvine’s lab, acts as an adjuvant that helps recruit a stronger B cell response to the vaccine.

In clinical trials, this vaccine and other experimental vaccines have been given as just one dose. However, there is growing evidence that a series of doses is more effective at generating broadly neutralizing antibodies. The seven-dose regimen, the researchers believe, works well because it mimics what happens when the body is exposed to a virus: The immune system builds up a strong response as more viral proteins, or antigens, accumulate in the body.

In the new study, the MIT team investigated how this response develops and explored whether they could achieve the same effect using a smaller number of vaccine doses.

“Giving seven doses just isn’t feasible for mass vaccination,” Bhagchandani says. “We wanted to identify some of the critical elements necessary for the success of this escalating dose, and to explore whether that knowledge could allow us to reduce the number of doses.”

The researchers began by comparing the effects of one, two, three, four, five, six, or seven doses, all given over a 12-day period. They initially found that while three or more doses generated strong antibody responses, two doses did not. However, by tweaking the dose intervals and ratios, the researchers discovered that giving 20 percent of the vaccine in the first dose and 80 percent in a second dose, seven days later, achieved just as good a response as the seven-dose schedule.

“It was clear that understanding the mechanisms behind this phenomenon would be crucial for future clinical translation,” Yang says. “Even if the ideal dosing ratio and timing may differ for humans, the underlying mechanistic principles will likely remain the same.”

Using a computational model, the researchers explored what was happening in each of these dosing scenarios. This work showed that when all of the vaccine is given as one dose, most of the antigen gets chopped into fragments before it reaches the lymph nodes. Lymph nodes are where B cells become activated to target a particular antigen, within structures known as germinal centers.

When only a tiny amount of the intact antigen reaches these germinal centers, B cells can’t come up with a strong response against that antigen.

However, a very small number of B cells do arise that produce antibodies targeting the intact antigen. So, giving a small amount in the first dose does not “waste” much antigen but allows some B cells and antibodies to develop. If a second, larger dose is given a week later, those antibodies bind to the antigen before it can be broken down and escort it into the lymph node. This allows more B cells to be exposed to that antigen and eventually leads to a large population of B cells that can target it.

“The early doses generate some small amounts of antibody, and that’s enough to then bind to the vaccine of the later doses, protect it, and target it to the lymph node. That's how we realized that we don't need to give seven doses,” Bhagchandani says. “A small initial dose will generate this antibody and then when you give the larger dose, it can again be protected because that antibody will bind to it and traffic it to the lymph node.”

T-cell boost

Those antigens may stay in the germinal centers for weeks or even longer, allowing more B cells to come in and be exposed to them, making it more likely that diverse types of antibodies will develop.

The researchers also found that the two-dose schedule induces a stronger T-cell response. The first dose activates dendritic cells, which promote inflammation and T-cell activation. Then, when the second dose arrives, even more dendritic cells are stimulated, further boosting the T-cell response.

Overall, the two-dose regimen resulted in a fivefold improvement in the T-cell response and a 60-fold improvement in the antibody response, compared to a single vaccine dose.

“Reducing the ‘escalating dose’ strategy down to two shots makes it much more practical for clinical implementation. Further, a number of technologies are in development that could mimic the two-dose exposure in a single shot, which could become ideal for mass vaccination campaigns,” Irvine says.

The researchers are now studying this vaccine strategy in a nonhuman primate model. They are also working on specialized materials that can deliver the second dose over an extended period of time, which could further enhance the immune response.

The research was funded by the Koch Institute Support (core) Grant from the National Cancer Institute, the National Institutes of Health, and the Ragon Institute of MIT, MGH, and Harvard.

What if construction materials could be put together and taken apart as easily as LEGO bricks? Such reconfigurable masonry would be disassembled at the end of a building’s lifetime and reassembled into a new structure, in a sustainable cycle that could supply generations of buildings using the same physical building blocks.

That’s the idea behind circular construction, which aims to reuse and repurpose a building’s materials whenever possible, to minimize the manufacturing of new materials and reduce the construction industry’s “embodied carbon,” which refers to the greenhouse gas emissions associated with every process throughout a building’s construction, from manufacturing to demolition.

Now MIT engineers, motivated by circular construction’s eco potential, are developing a new kind of reconfigurable masonry made from 3D-printed, recycled glass. Using a custom 3D glass printing technology provided by MIT spinoff Evenline, the team has made strong, multilayered glass bricks, each in the shape of a figure eight, that are designed to interlock, much like LEGO bricks.

In mechanical testing, a single glass brick withstood pressures similar to that of a concrete block. As a structural demonstration, the researchers constructed a wall of interlocking glass bricks. They envision that 3D-printable glass masonry could be reused many times over as recyclable bricks for building facades and internal walls.

“Glass is a highly recyclable material,” says Kaitlyn Becker, assistant professor of mechanical engineering at MIT. “We’re taking glass and turning it into masonry that, at the end of a structure’s life, can be disassembled and reassembled into a new structure, or can be stuck back into the printer and turned into a completely different shape. All this builds into our idea of a sustainable, circular building material.”

“Glass as a structural material kind of breaks people’s brains a little bit,” says Michael Stern, a former MIT graduate student and researcher in both MIT’s Media Lab and Lincoln Laboratory, who is also founder and director of Evenline. “We’re showing this is an opportunity to push the limits of what’s been done in architecture.”

Becker and Stern, with their colleagues, detail their glass brick design in a study appearing today in the journal Glass Structures and Engineering. Their MIT co-authors include lead author Daniel Massimino and Charlotte Folinus, along with Ethan Townsend at Evenline.

Lock step

The inspiration for the new circular masonry design arose partly in MIT’s Glass Lab, where Becker and Stern, then undergraduate students, first learned the art and science of blowing glass.

“I found the material fascinating,” says Stern, who later designed a 3D printer capable of printing molten recycled glass — a project he took on while studying in the mechanical engineering department. “I started thinking of how glass printing can find its place and do interesting things, construction being one possible route.”

Meanwhile, Becker, who accepted a faculty position at MIT, began exploring the intersection of manufacturing and design, and ways to develop new processes that enable innovative designs.

“I get excited about expanding design and manufacturing spaces for challenging materials with interesting characteristics, like glass and its optical properties and recyclability,” Becker says. “As long as it’s not contaminated, you can recycle glass almost infinitely.”

She and Stern teamed up to see whether and how 3D-printable glass could be made into a structural masonry unit as sturdy and stackable as traditional bricks. For their new study, the team used the Glass 3D Printer 3 (G3DP3), the latest version of Evenline’s glass printer, which pairs with a furnace to melt crushed glass bottles into a molten, printable form that the printer then deposits in layered patterns.

The team printed prototype glass bricks using soda-lime glass that is typically used in a glassblowing studio. They incorporated two round pegs onto each printed brick, similar to the studs on a LEGO brick. Like the toy blocks, the pegs enable bricks to interlock and assemble into larger structures. Another material placed between the bricks prevent scratches or cracks between glass surfaces but can be removed if a brick structure were to be dismantled and recycled, also allowing bricks to be remelted in the printer and formed into new shapes. The team decided to make the blocks into a figure-eight shape.

“With the figure-eight shape, we can constrain the bricks while also assembling them into walls that have some curvature,” Massimino says.

Stepping stones

The team printed glass bricks and tested their mechanical strength in an industrial hydraulic press that squeezed the bricks until they began to fracture. The researchers found that the strongest bricks were able to hold up to pressures that are comparable to what concrete blocks can withstand. Those strongest bricks were made mostly from printed glass, with a separately manufactured interlocking feature that attached to the bottom of the brick. These results suggest that most of a masonry brick could be made from printed glass, with an interlocking feature that could be printed, cast, or separately manufactured from a different material.

“Glass is a complicated material to work with,” Becker says. “The interlocking elements, made from a different material, showed the most promise at this stage.”

The group is looking into whether more of a brick’s interlocking feature could be made from printed glass, but doesn’t see this as a dealbreaker in moving forward to scale up the design. To demonstrate glass masonry’s potential, they constructed a curved wall of interlocking glass bricks. Next, they aim to build progressively bigger, self-supporting glass structures.

“We have more understanding of what the material’s limits are, and how to scale,” Stern says. “We’re thinking of stepping stones to buildings, and want to start with something like a pavilion — a temporary structure that humans can interact with, and that you could then reconfigure into a second design. And you could imagine that these blocks could go through a lot of lives.”

This research was supported, in part, by the Bose Research Grant Program and MIT’s Research Support Committee.

For more than 100 years, scientists have been using X-ray crystallography to determine the structure of crystalline materials such as metals, rocks, and ceramics.

This technique works best when the crystal is intact, but in many cases, scientists have only a powdered version of the material, which contains random fragments of the crystal. This makes it more challenging to piece together the overall structure.

MIT chemists have now come up with a new generative AI model that can make it much easier to determine the structures of these powdered crystals. The prediction model could help researchers characterize materials for use in batteries, magnets, and many other applications.

“Structure is the first thing that you need to know for any material. It’s important for superconductivity, it’s important for magnets, it’s important for knowing what photovoltaic you created. It’s important for any application that you can think of which is materials-centric,” says Danna Freedman, the Frederick George Keyes Professor of Chemistry at MIT.

Freedman and Jure Leskovec, a professor of computer science at Stanford University, are the senior authors of the new study, which appears today in the Journal of the American Chemical Society. MIT graduate student Eric Riesel and Yale University undergraduate Tsach Mackey are the lead authors of the paper.

Distinctive patterns

Crystalline materials, which include metals and most other inorganic solid materials, are made of lattices that consist of many identical, repeating units. These units can be thought of as “boxes” with a distinctive shape and size, with atoms arranged precisely within them.

When X-rays are beamed at these lattices, they diffract off atoms with different angles and intensities, revealing information about the positions of the atoms and the bonds between them. Since the early 1900s, this technique has been used to analyze materials, including biological molecules that have a crystalline structure, such as DNA and some proteins.

For materials that exist only as a powdered crystal, solving these structures becomes much more difficult because the fragments don’t carry the full 3D structure of the original crystal.

“The precise lattice still exists, because what we call a powder is really a collection of microcrystals. So, you have the same lattice as a large crystal, but they’re in a fully randomized orientation,” Freedman says.

For thousands of these materials, X-ray diffraction patterns exist but remain unsolved. To try to crack the structures of these materials, Freedman and her colleagues trained a machine-learning model on data from a database called the Materials Project, which contains more than 150,000 materials. First, they fed tens of thousands of these materials into an existing model that can simulate what the X-ray diffraction patterns would look like. Then, they used those patterns to train their AI model, which they call Crystalyze, to predict structures based on the X-ray patterns.

The model breaks the process of predicting structures into several subtasks. First, it determines the size and shape of the lattice “box” and which atoms will go into it. Then, it predicts the arrangement of atoms within the box. For each diffraction pattern, the model generates several possible structures, which can be tested by feeding the structures into a model that determines diffraction patterns for a given structure.

“Our model is generative AI, meaning that it generates something that it hasn’t seen before, and that allows us to generate several different guesses,” Riesel says. “We can make a hundred guesses, and then we can predict what the powder pattern should look like for our guesses. And then if the input looks exactly like the output, then we know we got it right.”

Solving unknown structures

The researchers tested the model on several thousand simulated diffraction patterns from the Materials Project. They also tested it on more than 100 experimental diffraction patterns from the RRUFF database, which contains powdered X-ray diffraction data for nearly 14,000 natural crystalline minerals, that they had held out of the training data. On these data, the model was accurate about 67 percent of the time. Then, they began testing the model on diffraction patterns that hadn’t been solved before. These data came from the Powder Diffraction File, which contains diffraction data for more than 400,000 solved and unsolved materials.

Using their model, the researchers came up with structures for more than 100 of these previously unsolved patterns. They also used their model to discover structures for three materials that Freedman’s lab created by forcing elements that do not react at atmospheric pressure to form compounds under high pressure. This approach can be used to generate new materials that have radically different crystal structures and physical properties, even though their chemical composition is the same.

Graphite and diamond — both made of pure carbon — are examples of such materials. The materials that Freedman has developed, which each contain bismuth and one other element, could be useful in the design of new materials for permanent magnets.

“We found a lot of new materials from existing data, and most importantly, solved three unknown structures from our lab that comprise the first new binary phases of those combinations of elements,” Freedman says.

Being able to determine the structures of powdered crystalline materials could help researchers working in nearly any materials-related field, according to the MIT team, which has posted a web interface for the model at crystalyze.org.

The research was funded by the U.S. Department of Energy and the National Science Foundation.

A new study from researchers at MIT and Penn State University reveals that if large language models were to be used in home surveillance, they could recommend calling the police even when surveillance videos show no criminal activity.

In addition, the models the researchers studied were inconsistent in which videos they flagged for police intervention. For instance, a model might flag one video that shows a vehicle break-in but not flag another video that shows a similar activity. Models often disagreed with one another over whether to call the police for the same video.

Furthermore, the researchers found that some models flagged videos for police intervention relatively less often in neighborhoods where most residents are white, controlling for other factors. This shows that the models exhibit inherent biases influenced by the demographics of a neighborhood, the researchers say.

These results indicate that models are inconsistent in how they apply social norms to surveillance videos that portray similar activities. This phenomenon, which the researchers call norm inconsistency, makes it difficult to predict how models would behave in different contexts.

“The move-fast, break-things modus operandi of deploying generative AI models everywhere, and particularly in high-stakes settings, deserves much more thought since it could be quite harmful,” says co-senior author Ashia Wilson, the Lister Brothers Career Development Professor in the Department of Electrical Engineering and Computer Science and a principal investigator in the Laboratory for Information and Decision Systems (LIDS).

Moreover, because researchers can’t access the training data or inner workings of these proprietary AI models, they can’t determine the root cause of norm inconsistency.

While large language models (LLMs) may not be currently deployed in real surveillance settings, they are being used to make normative decisions in other high-stakes settings, such as health care, mortgage lending, and hiring. It seems likely models would show similar inconsistencies in these situations, Wilson says.

“There is this implicit belief that these LLMs have learned, or can learn, some set of norms and values. Our work is showing that is not the case. Maybe all they are learning is arbitrary patterns or noise,” says lead author Shomik Jain, a graduate student in the Institute for Data, Systems, and Society (IDSS).

Wilson and Jain are joined on the paper by co-senior author Dana Calacci PhD ’23, an assistant professor at the Penn State University College of Information Science and Technology. The research will be presented at the AAAI Conference on AI, Ethics, and Society.

“A real, imminent, practical threat”

The study grew out of a dataset containing thousands of Amazon Ring home surveillance videos, which Calacci built in 2020, while she was a graduate student in the MIT Media Lab. Ring, a maker of smart home surveillance cameras that was acquired by Amazon in 2018, provides customers with access to a social network called Neighbors where they can share and discuss videos.

Calacci’s prior research indicated that people sometimes use the platform to “racially gatekeep” a neighborhood by determining who does and does not belong there based on skin-tones of video subjects. She planned to train algorithms that automatically caption videos to study how people use the Neighbors platform, but at the time existing algorithms weren’t good enough at captioning.

The project pivoted with the explosion of LLMs.

“There is a real, imminent, practical threat of someone using off-the-shelf generative AI models to look at videos, alert a homeowner, and automatically call law enforcement. We wanted to understand how risky that was,” Calacci says.

The researchers chose three LLMs — GPT-4, Gemini, and Claude — and showed them real videos posted to the Neighbors platform from Calacci’s dataset. They asked the models two questions: “Is a crime happening in the video?” and “Would the model recommend calling the police?”

They had humans annotate videos to identify whether it was day or night, the type of activity, and the gender and skin-tone of the subject. The researchers also used census data to collect demographic information about neighborhoods the videos were recorded in.

Inconsistent decisions

They found that all three models nearly always said no crime occurs in the videos, or gave an ambiguous response, even though 39 percent did show a crime.

“Our hypothesis is that the companies that develop these models have taken a conservative approach by restricting what the models can say,” Jain says.

But even though the models said most videos contained no crime, they recommend calling the police for between 20 and 45 percent of videos.

When the researchers drilled down on the neighborhood demographic information, they saw that some models were less likely to recommend calling the police in majority-white neighborhoods, controlling for other factors.

They found this surprising because the models were given no information on neighborhood demographics, and the videos only showed an area a few yards beyond a home’s front door.

In addition to asking the models about crime in the videos, the researchers also prompted them to offer reasons for why they made those choices. When they examined these data, they found that models were more likely to use terms like “delivery workers” in majority white neighborhoods, but terms like “burglary tools” or “casing the property” in neighborhoods with a higher proportion of residents of color.

“Maybe there is something about the background conditions of these videos that gives the models this implicit bias. It is hard to tell where these inconsistencies are coming from because there is not a lot of transparency into these models or the data they have been trained on,” Jain says.

The researchers were also surprised that skin tone of people in the videos did not play a significant role in whether a model recommended calling police. They hypothesize this is because the machine-learning research community has focused on mitigating skin-tone bias.

“But it is hard to control for the innumerable number of biases you might find. It is almost like a game of whack-a-mole. You can mitigate one and another bias pops up somewhere else,” Jain says.

Many mitigation techniques require knowing the bias at the outset. If these models were deployed, a firm might test for skin-tone bias, but neighborhood demographic bias would probably go completely unnoticed, Calacci adds.

“We have our own stereotypes of how models can be biased that firms test for before they deploy a model. Our results show that is not enough,” she says.

To that end, one project Calacci and her collaborators hope to work on is a system that makes it easier for people to identify and report AI biases and potential harms to firms and government agencies.

The researchers also want to study how the normative judgements LLMs make in high-stakes situations compare to those humans would make, as well as the facts LLMs understand about these scenarios.

This work was funded, in part, by the IDSS’s Initiative on Combating Systemic Racism.

When Jared Bryan talks about his seismology research, it’s with a natural finesse. He’s a fifth-year PhD student working with MIT Assistant Professor William Frank on seismology research, drawn in by the lab’s combination of GPS observations, satellites, and seismic station data to understand the underlying physics of earthquakes. He has no trouble talking about seismic velocity in fault zones or how he first became interested in the field after summer internships with the Southern California Earthquake Center as an undergraduate student.

“It’s definitely like a more down-to-earth kind of seismology,” he jokingly describes it. It’s an odd comment. Where else could earthquakes be but on Earth? But it’s because Bryan finished a research project that has culminated in a new paper — published today in Nature Astronomy — involving seismic activity not on Earth, but on stars.

Building curiosity

PhD students in MIT’s Department of Earth, Atmospheric and Planetary Sciences (EAPS) are required to complete two research projects as part of their general exam. The first is often in their main focus of research and the foundations of what will become their thesis work.

But the second project has a special requirement: It must be in a different specialty.

“Having that built into the structure of the PhD is really, really nice,” says Bryan, who hadn’t known about the special requirement when he decided to come to EAPS. “I think it helps you build curiosity and find what's interesting about what other people are doing.”

Having so many different, yet still related, fields of study housed in one department makes it easier for students with a strong sense of curiosity to explore the interconnected interactions of Earth science.

“I think everyone here is excited about a lot of different stuff, but we can’t do everything,” says Frank, the Victor P. Starr Career Development Professor of Geophysics. “This is a great way to get students to try something else that they maybe would have wanted to do in a parallel dimension, interact with other advisors, and see that science can be done in different ways.”

At first, Bryan was worried that the nature of the second project would be a restrictive diversion from his main PhD research. But Associate Professor Julien de Wit was looking for someone with a seismology background to look at some stellar observations he’d collected back in 2016. A star’s brightness was pulsating at a very specific frequency that had to be caused by changes in the star itself, so Bryan decided to help.

“I was surprised by how the kind of seismology that he was looking for was similar to the seismology that we were first doing in the ’60s and ’70s, like large-scale global Earth seismology,” says Bryan. “I thought it would be a way to rethink the foundations of the field that I had been studying applied to a new region.”

Going from earthquakes to starquakes is not a one-to-one comparison. While the foundational knowledge was there, movement of stars comes from a variety of sources like magnetism or the Coriolis effect, and in a variety of forms. In addition to the sound and pressure waves of earthquakes, they also have gravity waves, all of which happen on a scale much more massive.

“You have to stretch your mind a bit, because you can’t actually visit these places,” Bryan says. “It’s an unbelievable luxury that we have in Earth seismology that the things that we study are on Google Maps.”

But there are benefits to bringing in scientists from outside an area of expertise. De Wit, who served as Bryan’s supervisor for the project and is also an author on the paper, points out that they bring a fresh perspective and approach by asking unique questions.

“Things that people in the field would just take for granted are challenged by their questions,” he says, adding that Bryan was transparent about what he did and didn’t know, allowing for a rich exchange of information.

Tidal resonance locking

Bryan eventually found that the changes in the star’s brightness were caused by tidal resonance. Resonance is a physical occurrence where waves interact and amplify each other. The most common analogy is pushing someone on a swing set; when the person pushing does it at just the right time, it helps the person on the swing go higher.

“Tidal resonance is where you’re pushing at exactly the same frequency as they’re swinging, and the locking happens when both of those frequencies are changing,” Bryan explains. The person pushing the swing gets tired and pushes less often, while the chain of the swing change length. (Bryan jokes that here the analogy starts to break down.)

As a star changes over the course of its lifetime, tidal resonance locking can cause hot Jupiters, which are massive exoplanets that orbit very close to their host stars, to change orbital distances. This wandering migration, as they call it, explains how some hot Jupiters get so close to their host stars. They also found that the path they take to get there is not always smooth. It can speed up, slow down, or even regress.

An important implication from the paper is that tidal resonance locking could be used as an exoplanet detection tool, confirming de Wit’s hypothesis from the original 2016 observation that the pulsations had the potential to be used in such a way. If changes in the star’s brightness can be linked to this resonance locking, it may indicate planets that can’t be detected using current methods.

As below, so above

Most EAPS PhD students don’t advance their project beyond the requirements for the general exam, let alone get a paper out of it. At first, Bryan worried that continuing with it would end up being a distraction from his main work, but ultimately was glad that he committed to it and was able to contribute something meaningful to the emerging field of asteroseismology.

“I think it’s evidence that Jared is excited about what he does and has the drive and scientific skepticism to have done the extra steps to make sure that what he was doing was a real contribution to the scientific literature,” says Frank. “He’s a great example of success and what we hope for our students.”

While de Wit didn’t manage to convince Bryan to switch to exoplanet research permanently, he is “excited that there is the opportunity to keep on working together.”

Once he finishes his PhD, Bryan plans on continuing in academia as a professor running a research lab, shifting his focus onto volcano seismology and improving instrumentation for the field. He’s open to the possibility of taking his findings on Earth and applying them to volcanoes on other planetary bodies, such as those found on Venus and Jupiter’s moon Io.

“I’d like to be the bridge between those two things,” he says.

When Jared Bryan talks about his seismology research, it’s with a natural finesse. He’s a fifth-year PhD student working with MIT Assistant Professor William Frank on seismology research, drawn in by the lab’s combination of GPS observations, satellites, and seismic station data to understand the underlying physics of earthquakes. He has no trouble talking about seismic velocity in fault zones or how he first became interested in the field after summer internships with the Southern California Earthquake Center as an undergraduate student.

“It’s definitely like a more down-to-earth kind of seismology,” he jokingly describes it. It’s an odd comment. Where else could earthquakes be but on Earth? But it’s because Bryan finished a research project that has culminated in a new paper — published today in Nature Astronomy — involving seismic activity not on Earth, but on stars.

Building curiosity

PhD students in MIT’s Department of Earth, Atmospheric and Planetary Sciences (EAPS) are required to complete two research projects as part of their general exam. The first is often in their main focus of research and the foundations of what will become their thesis work.

But the second project has a special requirement: It must be in a different specialty.

“Having that built into the structure of the PhD is really, really nice,” says Bryan, who hadn’t known about the special requirement when he decided to come to EAPS. “I think it helps you build curiosity and find what's interesting about what other people are doing.”

Having so many different, yet still related, fields of study housed in one department makes it easier for students with a strong sense of curiosity to explore the interconnected interactions of Earth science.

“I think everyone here is excited about a lot of different stuff, but we can’t do everything,” says Frank, the Victor P. Starr Career Development Professor of Geophysics. “This is a great way to get students to try something else that they maybe would have wanted to do in a parallel dimension, interact with other advisors, and see that science can be done in different ways.”

At first, Bryan was worried that the nature of the second project would be a restrictive diversion from his main PhD research. But Associate Professor Julien de Wit was looking for someone with a seismology background to look at some stellar observations he’d collected back in 2016. A star’s brightness was pulsating at a very specific frequency that had to be caused by changes in the star itself, so Bryan decided to help.

“I was surprised by how the kind of seismology that he was looking for was similar to the seismology that we were first doing in the ’60s and ’70s, like large-scale global Earth seismology,” says Bryan. “I thought it would be a way to rethink the foundations of the field that I had been studying applied to a new region.”

Going from earthquakes to starquakes is not a one-to-one comparison. While the foundational knowledge was there, movement of stars comes from a variety of sources like magnetism or the Coriolis effect, and in a variety of forms. In addition to the sound and pressure waves of earthquakes, they also have gravity waves, all of which happen on a scale much more massive.

“You have to stretch your mind a bit, because you can’t actually visit these places,” Bryan says. “It’s an unbelievable luxury that we have in Earth seismology that the things that we study are on Google Maps.”

But there are benefits to bringing in scientists from outside an area of expertise. De Wit, who served as Bryan’s supervisor for the project and is also an author on the paper, points out that they bring a fresh perspective and approach by asking unique questions.

“Things that people in the field would just take for granted are challenged by their questions,” he says, adding that Bryan was transparent about what he did and didn’t know, allowing for a rich exchange of information.

Tidal resonance locking

Bryan eventually found that the changes in the star’s brightness were caused by tidal resonance. Resonance is a physical occurrence where waves interact and amplify each other. The most common analogy is pushing someone on a swing set; when the person pushing does it at just the right time, it helps the person on the swing go higher.

“Tidal resonance is where you’re pushing at exactly the same frequency as they’re swinging, and the locking happens when both of those frequencies are changing,” Bryan explains. The person pushing the swing gets tired and pushes less often, while the chain of the swing change length. (Bryan jokes that here the analogy starts to break down.)

As a star changes over the course of its lifetime, tidal resonance locking can cause hot Jupiters, which are massive exoplanets that orbit very close to their host stars, to change orbital distances. This wandering migration, as they call it, explains how some hot Jupiters get so close to their host stars. They also found that the path they take to get there is not always smooth. It can speed up, slow down, or even regress.

An important implication from the paper is that tidal resonance locking could be used as an exoplanet detection tool, confirming de Wit’s hypothesis from the original 2016 observation that the pulsations had the potential to be used in such a way. If changes in the star’s brightness can be linked to this resonance locking, it may indicate planets that can’t be detected using current methods.

As below, so above

Most EAPS PhD students don’t advance their project beyond the requirements for the general exam, let alone get a paper out of it. At first, Bryan worried that continuing with it would end up being a distraction from his main work, but ultimately was glad that he committed to it and was able to contribute something meaningful to the emerging field of asteroseismology.

“I think it’s evidence that Jared is excited about what he does and has the drive and scientific skepticism to have done the extra steps to make sure that what he was doing was a real contribution to the scientific literature,” says Frank. “He’s a great example of success and what we hope for our students.”

While de Wit didn’t manage to convince Bryan to switch to exoplanet research permanently, he is “excited that there is the opportunity to keep on working together.”

Once he finishes his PhD, Bryan plans on continuing in academia as a professor running a research lab, shifting his focus onto volcano seismology and improving instrumentation for the field. He’s open to the possibility of taking his findings on Earth and applying them to volcanoes on other planetary bodies, such as those found on Venus and Jupiter’s moon Io.

“I’d like to be the bridge between those two things,” he says.

In a new study, MIT physicists propose that if most of the dark matter in the universe is made up of microscopic primordial black holes — an idea first proposed in the 1970s — then these gravitational dwarfs should zoom through our solar system at least once per decade. A flyby like this, the researchers predict, would introduce a wobble into Mars’ orbit, to a degree that today’s technology could actually detect.

Such a detection could lend support to the idea that primordial black holes are a primary source of dark matter throughout the universe.

“Given decades of precision telemetry, scientists know the distance between Earth and Mars to an accuracy of about 10 centimeters,” says study author David Kaiser, professor of physics and the Germeshausen Professor of the History of Science at MIT. “We’re taking advantage of this highly instrumented region of space to try and look for a small effect. If we see it, that would count as a real reason to keep pursuing this delightful idea that all of dark matter consists of black holes that were spawned in less than a second after the Big Bang and have been streaming around the universe for 14 billion years.”

Kaiser and his colleagues report their findings today in the journal Physical Review D. The study’s co-authors are lead author Tung Tran ’24, who is now a graduate student at Stanford University; Sarah Geller ’12, SM ’17, PhD ’23, who is now a postdoc at the University of California at Santa Cruz; and MIT Pappalardo Fellow Benjamin Lehmann.

Beyond particles

Less than 20 percent of all physical matter is made from visible stuff, from stars and planets, to the kitchen sink. The rest is composed of dark matter, a hypothetical form of matter that is invisible across the entire electromagnetic spectrum yet is thought to pervade the universe and exert a gravitational force large enough to affect the motion of stars and galaxies.

Physicists have erected detectors on Earth to try and spot dark matter and pin down its properties. For the most part, these experiments assume that dark matter exists as a form of exotic particle that might scatter and decay into observable particles as it passes through a given experiment. But so far, such particle-based searches have come up empty.

In recent years, another possibility, first introduced in the 1970s, has regained traction: Rather than taking on a particle form, dark matter could exist as microscopic, primordial black holes that formed in the first moments following the Big Bang. Unlike the astrophysical black holes that form from the collapse of old stars, primordial black holes would have formed from the collapse of dense pockets of gas in the very early universe and would have scattered across the cosmos as the universe expanded and cooled.

These primordial black holes would have collapsed an enormous amount of mass into a tiny space. The majority of these primordial black holes could be as small as a single atom and as heavy as the largest asteroids. It would be conceivable, then, that such tiny giants could exert a gravitational force that could explain at least a portion of dark matter. For the MIT team, this possibility raised an initially frivolous question.

“I think someone asked me what would happen if a primordial black hole passed through a human body,” recalls Tung, who did a quick pencil-and-paper calculation to find that if such a black hole zinged within 1 meter of a person, the force of the black hole would push the person 6 meters, or about 20 feet away in a single second. Tung also found that the odds were astronomically unlikely that a primordial black hole would pass anywhere near a person on Earth.

Their interest piqued, the researchers took Tung’s calculations a step further, to estimate how a black hole flyby might affect much larger bodies such as the Earth and the moon.

“We extrapolated to see what would happen if a black hole flew by Earth and caused the moon to wobble by a little bit,” Tung says. “The numbers we got were not very clear. There are many other dynamics in the solar system that could act as some sort of friction to cause the wobble to dampen out.”

Close encounters

To get a clearer picture, the team generated a relatively simple simulation of the solar system that incorporates the orbits and gravitational interactions between all the planets, and some of the largest moons.

“State-of-the-art simulations of the solar system include more than a million objects, each of which has a tiny residual effect,” Lehmann notes. “But even modeling two dozen objects in a careful simulation, we could see there was a real effect that we could dig into.”

The team worked out the rate at which a primordial black hole should pass through the solar system, based on the amount of dark matter that is estimated to reside in a given region of space and the mass of a passing black hole, which in this case, they assumed to be as massive as the largest asteroids in the solar system, consistent with other astrophysical constraints.

“Primordial black holes do not live in the solar system. Rather, they’re streaming through the universe, doing their own thing,” says co-author Sarah Geller. “And the probability is, they’re going through the inner solar system at some angle once every 10 years or so.”

Given this rate, the researchers simulated various asteroid-mass black holes flying through the solar system, from various angles, and at velocities of about 150 miles per second. (The directions and speeds come from other studies of the distribution of dark matter throughout our galaxy.) They zeroed in on those flybys that appeared to be “close encounters,” or instances that caused some sort of effect in surrounding objects. They quickly found that any effect in the Earth or the moon was too uncertain to pin to a particular black hole. But Mars seemed to offer a clearer picture.

The researchers found that if a primordial black hole were to pass within a few hundred million miles of Mars, the encounter would set off a “wobble,” or a slight deviation in Mars’ orbit. Within a few years of such an encounter, Mars’ orbit should shift by about a meter — an incredibly small wobble, given the planet is more than 140 million miles from Earth. And yet, this wobble could be detected by the various high-precision instruments that are monitoring Mars today.

If such a wobble were detected in the next couple of decades, the researchers acknowledge there would still be much work needed to confirm that the push came from a passing black hole rather than a run-of-the-mill asteroid.

“We need as much clarity as we can of the expected backgrounds, such as the typical speeds and distributions of boring space rocks, versus these primordial black holes,” Kaiser notes. “Luckily for us, astronomers have been tracking ordinary space rocks for decades as they have flown through our solar system, so we could calculate typical properties of their trajectories and begin to compare them with the very different types of paths and speeds that primordial black holes should follow.”

To help with this, the researchers are exploring the possibility of a new collaboration with a group that has extensive expertise simulating many more objects in the solar system.

“We are now working to simulate a huge number of objects, from planets to moons and rocks, and how they’re all moving over long time scales,” Geller says. “We want to inject close encounter scenarios, and look at their effects with higher precision.”

“It’s a very neat test they’ve proposed, and it could tell us if the closest black hole is closer than we realize,” says Matt Caplan, associate professor of physics at Illinois State University, who was not involved in the study. “I should emphasize there’s a little bit of luck involved too. Whether or not a search finds a loud and clear signal depends on the exact path a wandering black hole takes through the solar system. Now that they’ve checked this idea with simulations, they have to do the hard part — checking the real data.”

This work was supported in part by the U.S. Department of Energy and the U.S. National Science Foundation, which includes an NSF Mathematical and Physical Sciences postdoctoral fellowship.

Ever been asked a question you only knew part of the answer to? To give a more informed response, your best move would be to phone a friend with more knowledge on the subject.

This collaborative process can also help large language models (LLMs) improve their accuracy. Still, it’s been difficult to teach LLMs to recognize when they should collaborate with another model on an answer. Instead of using complex formulas or large amounts of labeled data to spell out where models should work together, researchers at MIT’s Computer Science and Artificial Intelligence Laboratory (CSAIL) have envisioned a more organic approach.

Their new algorithm, called “Co-LLM,” can pair a general-purpose base LLM with a more specialized model and help them work together. As the former crafts an answer, Co-LLM reviews each word (or token) within its response to see where it can call upon a more accurate answer from the expert model. This process leads to more accurate replies to things like medical prompts and math and reasoning problems. Since the expert model is not needed at each iteration, this also leads to more efficient response generation.

To decide when a base model needs help from an expert model, the framework uses machine learning to train a “switch variable,” or a tool that can indicate the competence of each word within the two LLMs’ responses. The switch is like a project manager, finding areas where it should call in a specialist. If you asked Co-LLM to name some examples of extinct bear species, for instance, two models would draft answers together. The general-purpose LLM begins to put together a reply, with the switch variable intervening at the parts where it can slot in a better token from the expert model, such as adding the year when the bear species became extinct.

“With Co-LLM, we’re essentially training a general-purpose LLM to ‘phone’ an expert model when needed,” says Shannon Shen, an MIT PhD student in electrical engineering and computer science and CSAIL affiliate who’s a lead author on a new paper about the approach. “We use domain-specific data to teach the base model about its counterpart’s expertise in areas like biomedical tasks and math and reasoning questions. This process automatically finds the parts of the data that are hard for the base model to generate, and then it instructs the base model to switch to the expert LLM, which was pretrained on data from a similar field. The general-purpose model provides the ‘scaffolding’ generation, and when it calls on the specialized LLM, it prompts the expert to generate the desired tokens. Our findings indicate that the LLMs learn patterns of collaboration organically, resembling how humans recognize when to call upon an expert to fill in the blanks.”

A combination of flexibility and factuality

Imagine asking a general-purpose LLM to name the ingredients of a specific prescription drug. It may reply incorrectly, necessitating the expertise of a specialized model.

To showcase Co-LLM’s flexibility, the researchers used data like the BioASQ medical set to couple a base LLM with expert LLMs in different domains, like the Meditron model, which is pretrained on unlabeled medical data. This enabled the algorithm to help answer inquiries a biomedical expert would typically receive, such as naming the mechanisms causing a particular disease.

For example, if you asked a simple LLM alone to name the ingredients of a specific prescription drug, it may reply incorrectly. With the added expertise of a model that specializes in biomedical data, you’d get a more accurate answer. Co-LLM also alerts users where to double-check answers.

Another example of Co-LLM’s performance boost: When tasked with solving a math problem like “a3 · a2 if a=5,” the general-purpose model incorrectly calculated the answer to be 125. As Co-LLM trained the model to collaborate more with a large math LLM called Llemma, together they determined that the correct solution was 3,125.

Co-LLM gave more accurate replies than fine-tuned simple LLMs and untuned specialized models working independently. Co-LLM can guide two models that were trained differently to work together, whereas other effective LLM collaboration approaches, such as “Proxy Tuning,” need all of their component models to be trained similarly. Additionally, this baseline requires each model to be used simultaneously to produce the answer, whereas MIT’s algorithm simply activates its expert model for particular tokens, leading to more efficient generation.

When to ask the expert

The MIT researchers’ algorithm highlights that imitating human teamwork more closely can increase accuracy in multi-LLM collaboration. To further elevate its factual precision, the team may draw from human self-correction: They’re considering a more robust deferral approach that can backtrack when the expert model doesn’t give a correct response. This upgrade would allow Co-LLM to course-correct so the algorithm can still give a satisfactory reply.

The team would also like to update the expert model (via only training the base model) when new information is available, keeping answers as current as possible. This would allow Co-LLM to pair the most up-to-date information with strong reasoning power. Eventually, the model could assist with enterprise documents, using the latest information it has to update them accordingly. Co-LLM could also train small, private models to work with a more powerful LLM to improve documents that must remain within the server.

“Co-LLM presents an interesting approach for learning to choose between two models to improve efficiency and performance,” says Colin Raffel, associate professor at the University of Toronto and an associate research director at the Vector Institute, who wasn’t involved in the research. “Since routing decisions are made at the token-level, Co-LLM provides a granular way of deferring difficult generation steps to a more powerful model. The unique combination of model-token-level routing also provides a great deal of flexibility that similar methods lack. Co-LLM contributes to an important line of work that aims to develop ecosystems of specialized models to outperform expensive monolithic AI systems.”

Shen wrote the paper with four other CSAIL affiliates: PhD student Hunter Lang ’17, MEng ’18; former postdoc and Apple AI/ML researcher Bailin Wang; MIT assistant professor of electrical engineering and computer science Yoon Kim, and professor and Jameel Clinic member David Sontag PhD ’10, who are both part of MIT-IBM Watson AI Lab. Their research was supported, in part, by the National Science Foundation, The National Defense Science and Engineering Graduate (NDSEG) Fellowship, MIT-IBM Watson AI Lab, and Amazon. Their work was presented at the Annual Meeting of the Association for Computational Linguistics.

One of the brain’s most celebrated qualities is its adaptability. Changes to neural circuits, whose connections are continually adjusted as we experience and interact with the world, are key to how we learn. But to keep knowledge and memories intact, some parts of the circuitry must be resistant to this constant change.

“Brains have figured out how to navigate this landscape of balancing between stability and flexibility, so that you can have new learning and you can have lifelong memory,” says neuroscientist Mark Harnett, an investigator at MIT’s McGovern Institute for Brain Research. In the Aug. 27 issue of the journal Cell Reports, Harnett and his team show how individual neurons can contribute to both parts of this vital duality. By studying the synapses through which pyramidal neurons in the brain’s sensory cortex communicate, they have learned how the cells preserve their understanding of some of the world’s most fundamental features, while also maintaining the flexibility they need to adapt to a changing world.

Visual connections

Pyramidal neurons receive input from other neurons via thousands of connection points. Early in life, these synapses are extremely malleable; their strength can shift as a young animal takes in visual information and learns to interpret it. Most remain adaptable into adulthood, but Harnett’s team discovered that some of the cells’ synapses lose their flexibility when the animals are less than a month old. Having both stable and flexible synapses means these neurons can combine input from different sources to use visual information in flexible ways.

Postdoc Courtney Yaeger took a close look at these unusually stable synapses, which cluster together along a narrow region of the elaborately branched pyramidal cells. She was interested in the connections through which the cells receive primary visual information, so she traced their connections with neurons in a vision-processing center of the brain’s thalamus called the dorsal lateral geniculate nucleus (dLGN).

The long extensions through which a neuron receives signals from other cells are called dendrites, and they branch of from the main body of the cell into a tree-like structure. Spiny protrusions along the dendrites form the synapses that connect pyramidal neurons to other cells. Yaeger’s experiments showed that connections from the dLGN all led to a defined region of the pyramidal cells — a tight band within what she describes as the trunk of the dendritic tree.

Yaeger found several ways in which synapses in this region — formally known as the apical oblique dendrite domain — differ from other synapses on the same cells. “They’re not actually that far away from each other, but they have completely different properties,” she says.

Stable synapses

In one set of experiments, Yaeger activated synapses on the pyramidal neurons and measured the effect on the cells’ electrical potential. Changes to a neuron’s electrical potential generate the impulses the cells use to communicate with one another. It is common for a synapse’s electrical effects to amplify when synapses nearby are also activated. But when signals were delivered to the apical oblique dendrite domain, each one had the same effect, no matter how many synapses were stimulated. Synapses there don’t interact with one another at all, Harnett says. “They just do what they do. No matter what their neighbors are doing, they all just do kind of the same thing.”

The team was also able to visualize the molecular contents of individual synapses. This revealed a surprising lack of a certain kind of neurotransmitter receptor, called NMDA receptors, in the apical oblique dendrites. That was notable because of NMDA receptors’ role in mediating changes in the brain. “Generally when we think about any kind of learning and memory and plasticity, it’s NMDA receptors that do it,” Harnett says. “That is the by far most common substrate of learning and memory in all brains.”

When Yaeger stimulated the apical oblique synapses with electricity, generating patterns of activity that would strengthen most synapses, the team discovered a consequence of the limited presence of NMDA receptors. The synapses’ strength did not change. “There’s no activity-dependent plasticity going on there, as far as we have tested,” Yaeger says.

That makes sense, the researchers say, because the cells’ connections from the thalamus relay primary visual information detected by the eyes. It is through these connections that the brain learns to recognize basic visual features like shapes and lines.

“These synapses are basically a robust, high-fidelity readout of this visual information,” Harnett explains. “That’s what they’re conveying, and it’s not context-sensitive. So it doesn’t matter how many other synapses are active, they just do exactly what they’re going to do, and you can’t modify them up and down based on activity. So they’re very, very stable.”

“You actually don’t want those to be plastic,” adds Yaeger. "Can you imagine going to sleep and then forgetting what a vertical line looks like? That would be disastrous.” 

By conducting the same experiments in mice of different ages, the researchers determined that the synapses that connect pyramidal neurons to the thalamus become stable a few weeks after young mice first open their eyes. By that point, Harnett says, they have learned everything they need to learn. On the other hand, if mice spend the first weeks of their lives in the dark, the synapses never stabilize — further evidence that the transition depends on visual experience.

The team’s findings not only help explain how the brain balances flexibility and stability; they could help researchers teach artificial intelligence how to do the same thing. Harnett says artificial neural networks are notoriously bad at this: when an artificial neural network that does something well is trained to do something new, it almost always experiences “catastrophic forgetting” and can no longer perform its original task. Harnett’s team is exploring how they can use what they’ve learned about real brains to overcome this problem in artificial networks.

A collaboration between four MIT groups, led by principal investigators Laura L. Kiessling, Jeremiah A. Johnson, Alex K. Shalek, and Darrell J. Irvine, in conjunction with a group at Georgia Tech led by M.G. Finn, has revealed a new strategy for enabling immune system mobilization against cancer cells. The work, which appears today in ACS Nano, produces exactly the type of anti-tumor immunity needed to function as a tumor vaccine — both prophylactically and therapeutically.

Cancer cells can look very similar to the human cells from which they are derived. In contrast, viruses, bacteria, and fungi carry carbohydrates on their surfaces that are markedly different from those of human carbohydrates. Dendritic cells — the immune system’s best antigen-presenting cells — carry proteins on their surfaces that help them recognize these atypical carbohydrates and bring those antigens inside of them. The antigens are then processed into smaller peptides and presented to the immune system for a response. Intriguingly, some of these carbohydrate proteins can also collaborate to direct immune responses. This work presents a strategy for targeting those antigens to the dendritic cells that results in a more activated, stronger immune response.

Tackling tumors’ tenacity

The researchers’ new strategy shrouds the tumor antigens with foreign carbohydrates and co-delivers them with single-stranded RNA so that the dendritic cells can be programmed to recognize the tumor antigens as a potential threat. The researchers targeted the lectin (carbohydrate-binding protein) DC-SIGN because of its ability to serve as an activator of dendritic cell immunity. They decorated a virus-like particle (a particle composed of virus proteins assembled onto a piece of RNA that is noninfectious because its internal RNA is not from the virus) with DC-binding carbohydrate derivatives. The resulting glycan-costumed virus-like particles display unique sugars; therefore, the dendritic cells recognize them as something they need to attack.

“On the surface of the dendritic cells are carbohydrate binding proteins called lectins that combine to the sugars on the surface of bacteria or viruses, and when they do that they penetrate the membrane,” explains Kiessling, the paper’s senior author. “On the cell, the DC-SIGN gets clustered upon binding the virus or bacteria and that promotes internalization. When a virus-like particle gets internalized, it starts to fall apart and releases its RNA.” The toll-like receptor (bound to RNA) and DC-SIGN (bound to the sugar decoration) can both signal to activate the immune response.

Once the dendritic cells have sounded the alarm of a foreign invasion, a robust immune response is triggered that is significantly stronger than the immune response that would be expected with a typical untargeted vaccine. When an antigen is encountered by the dendritic cells, they send signals to T cells, the next cell in the immune system, to give different responses depending on what pathways have been activated in the dendritic cells.

Advancing cancer vaccine development

The activity of a potential vaccine developed in line with this new research is twofold. First, the vaccine glycan coat binds to lectins, providing a primary signal. Then, binding to toll-like receptors elicits potent immune activation.

The Kiessling, Finn, and Johnson groups had previously identified a synthetic DC-SIGN binding group that directed cellular immune responses when used to decorate virus-like particles. But it was unclear whether this method could be utilized as an anticancer vaccine. Collaboration between researchers in the labs at MIT and Georgia Tech demonstrated that in fact, it could.

Valerie Lensch, a chemistry PhD student from MIT’s Program in Polymers and Soft Matter and a joint member of the Kiessling and Johnson labs, took the preexisting strategy and tested it as an anticancer vaccine, learning a great deal about immunology in order to do so.

“We have developed a modular vaccine platform designed to drive antigen-specific cellular immune responses,” says Lensch. “This platform is not only pivotal in the fight against cancer, but also offers significant potential for combating challenging intracellular pathogens, including malaria parasites, HIV, and Mycobacterium tuberculosis. This technology holds promise for tackling a range of diseases where vaccine development has been particularly challenging.”

Lensch and her fellow researchers conducted in vitro experiments with extensive iterations of these glycan-costumed virus-like particles before identifying a design that demonstrated potential for success. Once that was achieved, the researchers were able to move on to an in vivo model, an exciting milestone for their research.

Adele Gabba, a postdoc in the Kiessling Lab, conducted the in vivo experiments with Lensch, and Robert Hincapie, who conducted his PhD studies with Professor M.G. Finn at Georgia Tech, built and decorated the virus-like particles with a series of glycans that were sent to him from the researchers at MIT.

“We are discovering that carbohydrates act like a language that cells use to communicate and direct the immune system,” says Gabba. “It's thrilling that we have begun to decode this language and can now harness it to reshape immune responses.”

“The design principles behind this vaccine are rooted in extensive fundamental research conducted by previous graduate student and postdoctoral researchers over many years, focusing on optimizing lectin engagement and understanding the roles of lectins in immunity,” says Lensch. “It has been exciting to witness the translation of these concepts into therapeutic platforms across various applications.”

A new study by MIT physicists proposes that a mysterious force known as early dark energy could solve two of the biggest puzzles in cosmology and fill in some major gaps in our understanding of how the early universe evolved.

One puzzle in question is the “Hubble tension,” which refers to a mismatch in measurements of how fast the universe is expanding. The other involves observations of numerous early, bright galaxies that existed at a time when the early universe should have been much less populated.

Now, the MIT team has found that both puzzles could be resolved if the early universe had one extra, fleeting ingredient: early dark energy. Dark energy is an unknown form of energy that physicists suspect is driving the expansion of the universe today. Early dark energy is a similar, hypothetical phenomenon that may have made only a brief appearance, influencing the expansion of the universe in its first moments before disappearing entirely.

Some physicists have suspected that early dark energy could be the key to solving the Hubble tension, as the mysterious force could accelerate the early expansion of the universe by an amount that would resolve the measurement mismatch.

The MIT researchers have now found that early dark energy could also explain the baffling number of bright galaxies that astronomers have observed in the early universe. In their new study, reported today in the Monthly Notices of the Royal Astronomical Society, the team modeled the formation of galaxies in the universe’s first few hundred million years. When they incorporated a dark energy component only in that earliest sliver of time, they found the number of galaxies that arose from the primordial environment bloomed to fit astronomers’ observations.

“You have these two looming open-ended puzzles,” says study co-author Rohan Naidu, a postdoc in MIT’s Kavli Institute for Astrophysics and Space Research. “We find that in fact, early dark energy is a very elegant and sparse solution to two of the most pressing problems in cosmology.”

The study’s co-authors include lead author and Kavli postdoc Xuejian (Jacob) Shen, and MIT professor of physics Mark Vogelsberger, along with Michael Boylan-Kolchin at the University of Texas at Austin, and Sandro Tacchella at the University of Cambridge.

Big city lights

Based on standard cosmological and galaxy formation models, the universe should have taken its time spinning up the first galaxies. It would have taken billions of years for primordial gas to coalesce into galaxies as large and bright as the Milky Way.

But in 2023, NASA’s James Webb Space Telescope (JWST) made a startling observation. With an ability to peer farther back in time than any observatory to date, the telescope uncovered a surprising number of bright galaxies as large as the modern Milky Way within the first 500 million years, when the universe was just 3 percent of its current age.

“The bright galaxies that JWST saw would be like seeing a clustering of lights around big cities, whereas theory predicts something like the light around more rural settings like Yellowstone National Park,” Shen says. “And we don’t expect that clustering of light so early on.”

For physicists, the observations imply that there is either something fundamentally wrong with the physics underlying the models or a missing ingredient in the early universe that scientists have not accounted for. The MIT team explored the possibility of the latter, and whether the missing ingredient might be early dark energy.

Physicists have proposed that early dark energy is a sort of antigravitational force that is turned on only at very early times. This force would counteract gravity’s inward pull and accelerate the early expansion of the universe, in a way that would resolve the mismatch in measurements. Early dark energy, therefore, is considered the most likely solution to the Hubble tension.

Galaxy skeleton

The MIT team explored whether early dark energy could also be the key to explaining the unexpected population of large, bright galaxies detected by JWST. In their new study, the physicists considered how early dark energy might affect the early structure of the universe that gave rise to the first galaxies. They focused on the formation of dark matter halos — regions of space where gravity happens to be stronger, and where matter begins to accumulate.

“We believe that dark matter halos are the invisible skeleton of the universe,” Shen explains. “Dark matter structures form first, and then galaxies form within these structures. So, we expect the number of bright galaxies should be proportional to the number of big dark matter halos.”

The team developed an empirical framework for early galaxy formation, which predicts the number, luminosity, and size of galaxies that should form in the early universe, given some measures of “cosmological parameters.” Cosmological parameters are the basic ingredients, or mathematical terms, that describe the evolution of the universe.

Physicists have determined that there are at least six main cosmological parameters, one of which is the Hubble constant — a term that describes the universe’s rate of expansion. Other parameters describe density fluctuations in the primordial soup, immediately after the Big Bang, from which dark matter halos eventually form.

The MIT team reasoned that if early dark energy affects the universe’s early expansion rate, in a way that resolves the Hubble tension, then it could affect the balance of the other cosmological parameters, in a way that might increase the number of bright galaxies that appear at early times. To test their theory, they incorporated a model of early dark energy (the same one that happens to resolve the Hubble tension) into an empirical galaxy formation framework to see how the earliest dark matter structures evolve and give rise to the first galaxies.

“What we show is, the skeletal structure of the early universe is altered in a subtle way where the amplitude of fluctuations goes up, and you get bigger halos, and brighter galaxies that are in place at earlier times, more so than in our more vanilla models,” Naidu says. “It means things were more abundant, and more clustered in the early universe.”

“A priori, I would not have expected the abundance of JWST’s early bright galaxies to have anything to do with early dark energy, but their observation that EDE pushes cosmological parameters in a direction that boosts the early-galaxy abundance is interesting,” says Marc Kamionkowski, professor of theoretical physics at Johns Hopkins University, who was not involved with the study. “I think more work will need to be done to establish a link between early galaxies and EDE, but regardless of how things turn out, it’s a clever — and hopefully ultimately fruitful — thing to try.”

“We demonstrated the potential of early dark energy as a unified solution to the two major issues faced by cosmology. This might be an evidence for its existence if the observational findings of JWST get further consolidated,” Vogelsberger concludes. “In the future, we can incorporate this into large cosmological simulations to see what detailed predictions we get.”

This research was supported, in part, by NASA and the National Science Foundation.

Placebos are inert treatments, generally not expected to impact biological pathways or improve a person’s physical health. But time and again, some patients report that they feel better after taking a placebo. Increasingly, doctors and scientists are recognizing that rather than dismissing placebos as mere trickery, they may be able to help patients by harnessing their power.

To maximize the impact of the placebo effect and design reliable therapeutic strategies, researchers need a better understanding of how it works. Now, with a new animal model developed by scientists at the McGovern Institute at MIT, they will be able to investigate the neural circuits that underlie placebos’ ability to elicit pain relief.

“The brain and body interaction has a lot of potential, in a way that we don't fully understand,” says Fan Wang, an MIT professor of brain and cognitive sciences and investigator at the McGovern Institute. “I really think there needs to be more of a push to understand placebo effect, in pain and probably in many other conditions. Now we have a strong model to probe the circuit mechanism.”

Context-dependent placebo effect

In the Sept. 5, 2024, issue of the journal Current Biology, Wang and her team report that they have elicited strong placebo pain relief in mice by activating pain-suppressing neurons in the brain while the mice are in a specific environment, thereby teaching the animals that they feel better when they are in that context. Following their training, placing the mice in that environment alone is enough to suppress pain. The team’s experiments — which were funded by the National Institutes of Health, the K. Lisa Yang Brain-Body Center, and the K. Lisa Yang and Hock E. Tan Center for Molecular Therapeutics within MIT’s Yang Tan Collective — show that this context-dependent placebo effect relieves both acute and chronic pain.

Context is critical for the placebo effect. While a pill can help a patient feel better when they expect it to, even if it is made only of sugar or starch, it seems to be not just the pill that sets up those expectations, but the entire scenario in which the pill is taken. For example, being in a hospital and interacting with doctors can contribute to a patient’s perception of care, and these social and environmental factors can make a placebo effect more probable.

MIT postdocs Bin Chen and Nitsan Goldstein used visual and textural cues to define a specific place. Then they activated pain-suppressing neurons in the brain while the animals were in this “pain-relief box.” Those pain-suppressing neurons, which Wang’s lab discovered a few years ago, are located in an emotion-processing center of the brain called the central amygdala. By expressing light-sensitive channels in these neurons, the researchers were able to suppress pain with light in the pain-relief box and leave the neurons inactive when mice were in a control box.

Animals learned to prefer the pain-relief box to other environments. And when the researchers tested their response to potentially painful stimuli after they had made that association, they found the mice were less sensitive while they were there. “Just by being in the context that they had associated with pain suppression, we saw that reduced pain—even though we weren't actually activating those [pain-suppressing] neurons,” Goldstein explains.

Acute and chronic pain relief

Some scientists have been able to elicit placebo pain relief in rodents by treating the animals with morphine, linking environmental cues to the pain suppression caused by the drugs similar to the way Wang’s team did by directly activating pain-suppressing neurons. This drug-based approach works best for setting up expectations of relief for acute pain; its placebo effect is short-lived and mostly ineffective against chronic pain. So Wang, Chen, and Goldstein were particularly pleased to find that their engineered placebo effect was effective for relieving both acute and chronic pain.

In their experiments, animals experiencing a chemotherapy-induced hypersensitivity to touch exhibited a preference for the pain relief box as much as animals who were exposed to a chemical that induces acute pain, days after their initial conditioning. Once there, their chemotherapy-induced pain sensitivity was eliminated; they exhibited no more sensitivity to painful stimuli than they had prior to receiving chemotherapy.

One of the biggest surprises came when the researchers turned their attention back to the pain-suppressing neurons in the central amygdala that they had used to trigger pain relief. They suspected that those neurons might be reactivated when mice returned to the pain-relief box. Instead, they found that after the initial conditioning period, those neurons remained quiet. “These neurons are not reactivated, yet the mice appear to be no longer in pain,” Wang says. “So it suggests this memory of feeling well is transferred somewhere else.”

Goldstein adds that there must be a pain-suppressing neural circuit somewhere that is activated by pain-relief-associated contexts — and the team’s new placebo model sets researchers up to investigate those pathways. A deeper understanding of that circuitry could enable clinicians to deploy the placebo effect — alone or in combination with active treatments — to better manage patients’ pain in the future.

To the untrained eye, a medical image like an MRI or X-ray appears to be a murky collection of black-and-white blobs. It can be a struggle to decipher where one structure (like a tumor) ends and another begins. 

When trained to understand the boundaries of biological structures, AI systems can segment (or delineate) regions of interest that doctors and biomedical workers want to monitor for diseases and other abnormalities. Instead of losing precious time tracing anatomy by hand across many images, an artificial assistant could do that for them.

The catch? Researchers and clinicians must label countless images to train their AI system before it can accurately segment. For example, you’d need to annotate the cerebral cortex in numerous MRI scans to train a supervised model to understand how the cortex’s shape can vary in different brains.

Sidestepping such tedious data collection, researchers from MIT’s Computer Science and Artificial Intelligence Laboratory (CSAIL), Massachusetts General Hospital (MGH), and Harvard Medical School have developed the interactive “ScribblePrompt” framework: a flexible tool that can help rapidly segment any medical image, even types it hasn’t seen before. 

Instead of having humans mark up each picture manually, the team simulated how users would annotate over 50,000 scans, including MRIs, ultrasounds, and photographs, across structures in the eyes, cells, brains, bones, skin, and more. To label all those scans, the team used algorithms to simulate how humans would scribble and click on different regions in medical images. In addition to commonly labeled regions, the team also used superpixel algorithms, which find parts of the image with similar values, to identify potential new regions of interest to medical researchers and train ScribblePrompt to segment them. This synthetic data prepared ScribblePrompt to handle real-world segmentation requests from users.

“AI has significant potential in analyzing images and other high-dimensional data to help humans do things more productively,” says MIT PhD student Hallee Wong SM ’22, the lead author on a new paper about ScribblePrompt and a CSAIL affiliate. “We want to augment, not replace, the efforts of medical workers through an interactive system. ScribblePrompt is a simple model with the efficiency to help doctors focus on the more interesting parts of their analysis. It’s faster and more accurate than comparable interactive segmentation methods, reducing annotation time by 28 percent compared to Meta’s Segment Anything Model (SAM) framework, for example.”

ScribblePrompt’s interface is simple: Users can scribble across the rough area they’d like segmented, or click on it, and the tool will highlight the entire structure or background as requested. For example, you can click on individual veins within a retinal (eye) scan. ScribblePrompt can also mark up a structure given a bounding box.

Then, the tool can make corrections based on the user’s feedback. If you wanted to highlight a kidney in an ultrasound, you could use a bounding box, and then scribble in additional parts of the structure if ScribblePrompt missed any edges. If you wanted to edit your segment, you could use a “negative scribble” to exclude certain regions.

These self-correcting, interactive capabilities made ScribblePrompt the preferred tool among neuroimaging researchers at MGH in a user study. 93.8 percent of these users favored the MIT approach over the SAM baseline in improving its segments in response to scribble corrections. As for click-based edits, 87.5 percent of the medical researchers preferred ScribblePrompt.

ScribblePrompt was trained on simulated scribbles and clicks on 54,000 images across 65 datasets, featuring scans of the eyes, thorax, spine, cells, skin, abdominal muscles, neck, brain, bones, teeth, and lesions. The model familiarized itself with 16 types of medical images, including microscopies, CT scans, X-rays, MRIs, ultrasounds, and photographs.

“Many existing methods don't respond well when users scribble across images because it’s hard to simulate such interactions in training. For ScribblePrompt, we were able to force our model to pay attention to different inputs using our synthetic segmentation tasks,” says Wong. “We wanted to train what’s essentially a foundation model on a lot of diverse data so it would generalize to new types of images and tasks.”

After taking in so much data, the team evaluated ScribblePrompt across 12 new datasets. Although it hadn’t seen these images before, it outperformed four existing methods by segmenting more efficiently and giving more accurate predictions about the exact regions users wanted highlighted.

“​​Segmentation is the most prevalent biomedical image analysis task, performed widely both in routine clinical practice and in research — which leads to it being both very diverse and a crucial, impactful step,” says senior author Adrian Dalca SM ’12, PhD ’16, CSAIL research scientist and assistant professor at MGH and Harvard Medical School. “ScribblePrompt was carefully designed to be practically useful to clinicians and researchers, and hence to substantially make this step much, much faster.”

“The majority of segmentation algorithms that have been developed in image analysis and machine learning are at least to some extent based on our ability to manually annotate images,” says Harvard Medical School professor in radiology and MGH neuroscientist Bruce Fischl, who was not involved in the paper. “The problem is dramatically worse in medical imaging in which our ‘images’ are typically 3D volumes, as human beings have no evolutionary or phenomenological reason to have any competency in annotating 3D images. ScribblePrompt enables manual annotation to be carried out much, much faster and more accurately, by training a network on precisely the types of interactions a human would typically have with an image while manually annotating. The result is an intuitive interface that allows annotators to naturally interact with imaging data with far greater productivity than was previously possible.”

Wong and Dalca wrote the paper with two other CSAIL affiliates: John Guttag, the Dugald C. Jackson Professor of EECS at MIT and CSAIL principal investigator; and MIT PhD student Marianne Rakic SM ’22. Their work was supported, in part, by Quanta Computer Inc., the Eric and Wendy Schmidt Center at the Broad Institute, the Wistron Corp., and the National Institute of Biomedical Imaging and Bioengineering of the National Institutes of Health, with hardware support from the Massachusetts Life Sciences Center.

Wong and her colleagues’ work will be presented at the 2024 European Conference on Computer Vision and was presented as an oral talk at the DCAMI workshop at the Computer Vision and Pattern Recognition Conference earlier this year. They were awarded the Bench-to-Bedside Paper Award at the workshop for ScribblePrompt’s potential clinical impact.

Sarah Sterling, director of the Cryo-Electron Microscopy, or Cryo-EM, core facility, often compares her job to running a small business. Each day brings a unique set of jobs ranging from administrative duties and managing facility users to balancing budgets and maintaining equipment.

Although one could easily be overwhelmed by the seemingly never-ending to-do list, Sterling finds a great deal of joy in wearing so many different hats. One of her most essential tasks involves clear communication with users when the delicate instruments in the facility are unusable because of routine maintenance and repairs.

“Better planning allows for better science,” Sterling says. “Luckily, I’m very comfortable with building and fixing things. Let’s troubleshoot. Let’s take it apart. Let’s put it back together.”

Out of all her duties as a core facility director, she most looks forward to the opportunities to teach, especially helping students develop research projects.

“Undergraduate or early-stage graduate students ask the best questions,” she says. “They’re so curious about the tiny details, and they’re always ready to hit the ground running on their projects.”

A non-linear scientific journey

When Sterling enrolled in Russell Sage College, a women’s college in New York, she was planning to pursue a career as a physical therapist. However, she quickly realized she loved her chemistry classes more than her other subjects. She graduated with a bachelor of science degree in chemistry and immediately enrolled in a master’s degree program in chemical engineering at the University of Maine.

Sterling was convinced to continue her studies at the University of Maine with a dual PhD in chemical engineering and biomedical sciences. That decision required the daunting process of taking two sets of core courses and completing a qualifying exam in each field. 

“I wouldn’t recommend doing that,” she says with a laugh. “To celebrate after finishing that intense experience, I took a year off to figure out what came next.”

Sterling chose to do a postdoc in the lab of Eva Nogales, a structural biology professor at the University of California at Berkeley. Nogales was looking for a scientist with experience working with lipids, a class of molecules that Sterling had studied extensively in graduate school.

At the time Sterling joined, the Nogales Lab was at the forefront of implementing an exciting structural biology approach: cryo-EM.

“When I was interviewing, I’d never even seen the type of microscope required for cryo-EM, let alone performed any experiments,” Sterling says. “But I remember thinking ‘I’m sure I can figure this out.’”

Cryo-EM is a technique that allows researchers to determine the three-dimensional shape, or structure, of the macromolecules that make up cells. A researcher can take a sample of their macromolecule of choice, suspend it in a liquid solution, and rapidly freeze it onto a grid to capture the macromolecules in random positions — the “cryo” part of the name. Powerful electron microscopes then collect images of the macromolecule — the EM part of cryo-EM. 

The two-dimensional images of the macromolecules from different angles can be combined to produce a three-dimensional structure. Structural information like this can reveal the macromolecule’s function inside cells or inform how it differs in a disease state. The rapidly expanding use of cryo-EM has unlocked so many mechanistic insights that the researchers who developed the technology were awarded the 2017 Nobel Prize in Chemistry. 

The MIT.nano facility opened its doors in 2018. The open-access, state-of-the-art facility now has more than 160 tools and more than 1,500 users representing nearly every department at MIT. The Cryo-EM facility lives in the basement of the MIT.nano building and houses multiple electron microscopes and laboratory space for cryo-specimen preparation.

Thanks to her work at UC Berkeley, Sterling’s career trajectory has long been intertwined with the expanding use of cryo-EM in research. Sterling anticipated the need for experienced scientists to run core facilities in order to maintain the electron microscopes needed for cryo-EM, which range in cost from a staggering $1 million to $10 million each.

After completing her postdoc, Sterling worked at the Harvard University cryo-EM core facility for five years. When the director position for the MIT.nano Cryo-EM facility opened, she decided to apply.

“I like that the core facility at MIT was smaller and more frequently used by students,” Sterling says. “There’s a lot more teaching, which is a challenge sometimes, but it’s rewarding to impact someone’s career at such an early stage.”

A focus on users

When Sterling arrived at MIT, her first initiative was to meet directly with all the students in research labs that use the core facility to learn what would make using the facility a better experience. She also implemented clear and standard operating procedures for cryo-EM beginners.

“I think being consistent and available has really improved users’ experiences,” Sterling says.

The users themselves report that her initiatives have proven highly successful — and have helped them grow as scientists.

“Sterling cultivates an environment where I can freely ask questions about anything to support my learning,” says Bonnie Su, a frequent Cryo-EM facility user and graduate student from the Vos lab.

But Sterling does not want to stop there. Looking ahead, she hopes to expand the facility by acquiring an additional electron microscope to allow more users to utilize this powerful technology in their research. She also plans to build a more collaborative community of cryo-EM scientists at MIT with additional symposia and casual interactions such as coffee hours.

Under her management, cryo-EM research has flourished. In the last year, the Cryo-EM core facility has supported research resulting in 12 new publications across five different departments at MIT. The facility has also provided access to 16 industry and non-MIT academic entities. These studies have revealed important insights into various biological processes, from visualizing how large protein machinery reads our DNA to the protein aggregates found in neurodegenerative disorders.

If anyone wants to conduct cryo-EM experiments or learn more about the technique, Sterling encourages anyone in the MIT community to reach out.

“Come visit us!” she says. “We give lots of tours, and you can stop by to say hi anytime.”

The idea of electrically stimulating a brain region called the central thalamus has gained traction among researchers and clinicians because it can help arouse subjects from unconscious states induced by traumatic brain injury or anesthesia, and can boost cognition and performance in awake animals. But the method, called CT-DBS, can have a side effect: seizures. A new study by researchers at MIT and Massachusetts General Hospital (MGH) who were testing the method in awake mice quantifies the probability of seizures at different stimulation currents and cautions that they sometimes occurred even at low levels.

“Understanding production and prevalence of this type of seizure activity is important because brain stimulation-based therapies are becoming more widely used,” says co-senior author Emery N. Brown, Edward Hood Taplin Professor of Medical Engineering and Computational Neuroscience in The Picower Institute for Learning and Memory, the Institute for Medical Engineering and Science, the Department of Brain and Cognitive Sciences, and the Center for Brains Minds and Machines (CBMM) at MIT.

In the brain, the seizures associated with CT-DBS occur as “electrographic seizures,” which are bursts of voltage among neurons across a broad spectrum of frequencies. Behaviorally, they manifest as “absence seizures” in which the subject appears to take on a blank stare and freezes for about 10-20 seconds.

In their study, the researchers were hoping to determine a CT-DBS stimulation current — in a clinically relevant range of under 200 microamps — below which seizures could be reliably avoided.

In search of that ideal current, they developed a protocol of starting brief bouts of CT-DBS at 1 microamp and then incrementally ramping the current up to 200 microamps until they found a threshold where an electrographic seizure occurred. Once they found that threshold, then they tested a longer bout of stimulation at the next lowest current level in hopes that an electrographic seizure wouldn’t occur. They did this for a variety of different stimulation frequencies. To their surprise, electrographic seizures still occurred 2.2 percent of the time during those longer stimulation trials (i.e. 22 times out of 996 tests) and in 10 out of 12 mice. At just 20 microamps, mice still experienced seizures in three out of 244 tests, a 1.2 percent rate.

“This is something that we needed to report because this was really surprising,” says co-lead author Francisco Flores, a research affiliate in The Picower Institute and CBMM, and an instructor in anesthesiology at MGH, where Brown is also an anesthesiologist. Isabella Dalla Betta, a technical associate in The Picower Institute, co-led the study published in Brain Stimulation.

Stimulation frequency didn’t matter for seizure risk but the rate of electrographic seizures increased as the current level increased. For instance, it happened in 5 out of 190 tests at 50 microamps, and two out of 65 tests at 100 microamps. The researchers also found that when an electrographic seizure occurred, it did so more quickly at higher currents than at lower levels. Finally, they also saw that seizures happened more quickly if they stimulated the thalamus on both sides of the brain, versus just one side. Some mice exhibited behaviors similar to absence seizure, though others became hyperactive.

It is not clear why some mice experienced electrographic seizures at just 20 microamps while two mice did not experience the seizures even at 200. Flores speculated that there may be different brain states that change the predisposition to seizures amid stimulation of the thalamus. Notably, seizures are not typically observed in humans who receive CT-DBS while in a minimally conscious state after a traumatic brain injury or in animals who are under anesthesia. Flores said the next stage of the research would aim to discern what the relevant brain states may be.

In the meantime, the study authors wrote, “EEG should be closely monitored for electrographic seizures when performing CT-DBS, especially in awake subjects.”

The paper’s co-senior author is Matt Wilson, Sherman Fairchild Professor in The Picower Institute, CBMM, and the departments of Biology and Brain and Cognitive Sciences. In addition to Dalla Betta, Flores, Brown and Wilson, the study’s other authors are John Tauber, David Schreier, and Emily Stephen.

Support for the research came from The JPB Foundation, The Picower Institute for Learning and Memory; George J. Elbaum ’59, SM ’63, PhD ’67, Mimi Jensen, Diane B. Greene SM ’78, Mendel Rosenblum, Bill Swanson, annual donors to the Anesthesia Initiative Fund; and the National Institutes of Health.

Typically, electrons are free agents that can move through most metals in any direction. When they encounter an obstacle, the charged particles experience friction and scatter randomly like colliding billiard balls.

But in certain exotic materials, electrons can appear to flow with single-minded purpose. In these materials, electrons may become locked to the material’s edge and flow in one direction, like ants marching single-file along a blanket’s boundary. In this rare “edge state,” electrons can flow without friction, gliding effortlessly around obstacles as they stick to their perimeter-focused flow. Unlike in a superconductor, where all electrons in a material flow without resistance, the current carried by edge modes occurs only at a material’s boundary.

Now MIT physicists have directly observed edge states in a cloud of ultracold atoms. For the first time, the team has captured images of atoms flowing along a boundary without resistance, even as obstacles are placed in their path. The results, which appear today in Nature Physics, could help physicists manipulate electrons to flow without friction in materials that could enable super-efficient, lossless transmission of energy and data.

“You could imagine making little pieces of a suitable material and putting it inside future devices, so electrons could shuttle along the edges and between different parts of your circuit without any loss,” says study co-author Richard Fletcher, assistant professor of physics at MIT. “I would stress though that, for us, the beauty is seeing with your own eyes physics which is absolutely incredible but usually hidden away in materials and unable to be viewed directly.”

The study’s co-authors at MIT include graduate students Ruixiao Yao and Sungjae Chi, former graduate students Biswaroop Mukherjee PhD ’20 and Airlia Shaffer PhD ’23, along with Martin Zwierlein, the Thomas A. Frank Professor of Physics. The co-authors are all members of MIT’s Research Laboratory of Electronics and the MIT-Harvard Center for Ultracold Atoms.

Forever on the edge

Physicists first invoked the idea of edge states to explain a curious phenomenon, known today as the Quantum Hall effect, which scientists first observed in 1980, in experiments with layered materials, where electrons were confined to two dimensions. These experiments were performed in ultracold conditions, and under a magnetic field. When scientists tried to send a current through these materials, they observed that electrons did not flow straight through the material, but instead accumulated on one side, in precise quantum portions.

To try and explain this strange phenomenon, physicists came up with the idea that these Hall currents are carried by edge states. They proposed that, under a magnetic field, electrons in an applied current could be deflected to the edges of a material, where they would flow and accumulate in a way that might explain the initial observations.

“The way charge flows under a magnetic field suggests there must be edge modes,” Fletcher says. “But to actually see them is quite a special thing because these states occur over femtoseconds, and across fractions of a nanometer, which is incredibly difficult to capture.”

Rather than try and catch electrons in an edge state, Fletcher and his colleagues realized they might be able to recreate the same physics in a larger and more observable system. The team has been studying the behavior of ultracold atoms in a carefully designed setup that mimics the physics of electrons under a magnetic field.

“In our setup, the same physics occurs in atoms, but over milliseconds and microns,” Zwierlein explains. “That means that we can take images and watch the atoms crawl essentially forever along the edge of the system.”

A spinning world

In their new study, the team worked with a cloud of about 1 million sodium atoms, which they corralled in a laser-controlled trap, and cooled to nanokelvin temperatures. They then manipulated the trap to spin the atoms around, much like riders on an amusement park Gravitron.

“The trap is trying to pull the atoms inward, but there’s centrifugal force that tries to pull them outward,” Fletcher explains. “The two forces balance each other, so if you’re an atom, you think you’re living in a flat space, even though your world is spinning. There’s also a third force, the Coriolis effect, such that if they try to move in a line, they get deflected. So these massive atoms now behave as if they were electrons living in a magnetic field.”

Into this manufactured reality, the researchers then introduced an “edge,” in the form of a ring of laser light, which formed a circular wall around the spinning atoms. As the team took images of the system, they observed that when the atoms encountered the ring of light, they flowed along its edge, in just one direction.

“You can imagine these are like marbles that you’ve spun up really fast in a bowl, and they just keep going around and around the rim of the bowl,” Zwierlein offers. “There is no friction. There is no slowing down, and no atoms leaking or scattering into the rest of the system. There is just beautiful, coherent flow.”

“These atoms are flowing, free of friction, for hundreds of microns,” Fletcher adds. “To flow that long, without any scattering, is a type of physics you don’t normally see in ultracold atom systems.”

This effortless flow held up even when the researchers placed an obstacle in the atoms’ path, like a speed bump, in the form of a point of light, which they shone along the edge of the original laser ring. Even as they came upon this new obstacle, the atoms didn’t slow their flow or scatter away, but instead glided right past without feeling friction as they normally would.

“We intentionally send in this big, repulsive green blob, and the atoms should bounce off it,” Fletcher says. “But instead what you see is that they magically find their way around it, go back to the wall, and continue on their merry way.”

The team’s observations in atoms document the same behavior that has been predicted to occur in electrons. Their results show that the setup of atoms is a reliable stand-in for studying how electrons would behave in edge states.

“It’s a very clean realization of a very beautiful piece of physics, and we can directly demonstrate the importance and reality of this edge,” Fletcher says. “A natural direction is to now introduce more obstacles and interactions into the system, where things become more unclear as to what to expect.”

This research was supported, in part, by the National Science Foundation.

Water contamination by the chemicals used in today’s technology is a rapidly growing problem globally. A recent study by the U.S. Centers for Disease Control found that 98 percent of people tested had detectable levels of PFAS, a family of particularly long-lasting compounds also known as “forever chemicals,” in their bloodstream.

A new filtration material developed by researchers at MIT might provide a nature-based solution to this stubborn contamination issue. The material, based on natural silk and cellulose, can remove a wide variety of these persistent chemicals as well as heavy metals. And, its antimicrobial properties can help keep the filters from fouling.

The findings are described in the journal ACS Nano, in a paper by MIT postdoc Yilin Zhang, professor of civil and environmental engineering Benedetto Marelli, and four others from MIT.

PFAS chemicals are present in a wide range of products, including cosmetics, food packaging, water-resistant clothing, firefighting foams, and antistick coating for cookware. A recent study identified 57,000 sites contaminated by these chemicals in the U.S. alone. The U.S. Environmental Protection Agency has estimated that PFAS remediation will cost $1.5 billion per year, in order to meet new regulations that call for limiting the compound to less than 7 parts per trillion in drinking water.

Contamination by PFAS and similar compounds “is actually a very big deal, and current solutions may only partially resolve this problem very efficiently or economically,” Zhang says. “That’s why we came up with this protein and cellulose-based, fully natural solution,” he says.

“We came to the project by chance,” Marelli notes. The initial technology that made the filtration material possible was developed by his group for a completely unrelated purpose — as a way to make a labelling system to counter the spread of counterfeit seeds, which are often of inferior quality. His team devised a way of processing silk proteins into uniform nanoscale crystals, or “nanofibrils,” through an environmentally benign, water-based drop-casting method at room temperature.

Zhang suggested that their new nanofibrillar material might be effective at filtering contaminants, but initial attempts with the silk nanofibrils alone didn’t work. The team decided to try adding another material: cellulose, which is abundantly available and can be obtained from agricultural wood pulp waste. The researchers used a self-assembly method in which the silk fibroin protein is suspended in water and then templated into nanofibrils by inserting “seeds” of cellulose nanocrystals. This causes the previously disordered silk molecules to line up together along the seeds, forming the basis of a hybrid material with distinct new properties.

By integrating cellulose into the silk-based fibrils that could be formed into a thin membrane, and then tuning the electrical charge of the cellulose, the researchers produced a material that was highly effective at removing contaminants in lab tests.

The electrical charge of the cellulose, they found, also gave it strong antimicrobial properties. This is a significant advantage, since one of the primary causes of failure in filtration membranes is fouling by bacteria and fungi. The antimicrobial properties of this material should greatly reduce that fouling issue, the researchers say.

“These materials can really compete with the current standard materials in water filtration when it comes to extracting metal ions and these emerging contaminants, and they can also outperform some of them currently,” Marelli says. In lab tests, the materials were able to extract orders of magnitude more of the contaminants from water than the currently used standard materials, activated carbon or granular activated carbon.

While the new work serves as a proof of principle, Marelli says, the team plans to continue working on improving the material, especially in terms of durability and availability of source materials. While the silk proteins used can be available as a byproduct of the silk textile industry, if this material were to be scaled up to address the global needs for water filtration, the supply might be insufficient. Also, alternative protein materials may turn out to perform the same function at lower cost.

Initially, the material would likely be used as a point-of-use filter, something that could be attached to a kitchen faucet, Zhang says. Eventually, it could be scaled up to provide filtration for municipal water supplies, but only after testing demonstrates that this would not pose any risk of introducing any contamination into the water supply. But one big advantage of the material, he says, is that both the silk and the cellulose constituents are considered food-grade substances, so any contamination is unlikely.

“Most of the normal materials available today are focusing on one class of contaminants or solving single problems,” Zhang says. “I think we are among the first to address all of these simultaneously.”

“What I love about this approach is that it is using only naturally grown materials like silk and cellulose to fight pollution,” says Hannes Schniepp, professor of applied science at the College of William and Mary, who was not associated with this work. “In competing approaches, synthetic materials are used — which usually require only more chemistry to fight some of the adverse outcomes that chemistry has produced. [This work] breaks this cycle! ... If this can be mass-produced in an economically viable way, this could really have a major impact.”

The research team included MIT postdocs Hui Sun and Meng Li, graduate student Maxwell Kalinowski, and recent graduate Yunteng Cao PhD ’22, now a postdoc at Yale University. The work was supported by the U.S. Office of Naval Research, the U.S. National Science Foundation, and the Singapore-MIT Alliance for Research and Technology.

Imagine how a phone call works: Your voice is converted into electronic signals, shifted up to higher frequencies, transmitted over long distances, and then shifted back down so it can be heard clearly on the other end. The process enabling this shifting of signal frequencies is called frequency mixing, and it is essential for communication technologies like radio and Wi-Fi. Frequency mixers are vital components in many electronic devices and typically operate using frequencies that oscillate billions (GHz, gigahertz) to trillions (THz, terahertz) of times per second. 

Now imagine a frequency mixer that works at a quadrillion (PHz, petahertz) times per second — up to a million times faster. This frequency range corresponds to the oscillations of the electric and magnetic fields that make up light waves. Petahertz-frequency mixers would allow us to shift signals up to optical frequencies and then back down to more conventional electronic frequencies, enabling the transmission and processing of vastly larger amounts of information at many times higher speeds. This leap in speed isn’t just about doing things faster; it’s about enabling entirely new capabilities.

Lightwave electronics (or petahertz electronics) is an emerging field that aims to integrate optical and electronic systems at incredibly high speeds, leveraging the ultrafast oscillations of light fields. The key idea is to harness the electric field of light waves, which oscillate on sub-femtosecond (10^-15 seconds) timescales, to directly drive electronic processes. This allows for the processing and manipulation of information at speeds far beyond what is possible with current electronic technologies. In combination with other petahertz electronic circuitry, a petahertz electronic mixer would allow us to process and analyze vast amounts of information in real time and transfer larger amounts of data over the air at unprecedented speeds. The MIT team’s demonstration of a lightwave-electronic mixer at petahertz-scale frequencies is a first step toward making communication technology faster, and progresses research toward developing new, miniaturized lightwave electronic circuitry capable of handling optical signals directly at the nanoscale.

In the 1970s, scientists began exploring ways to extend electronic frequency mixing into the terahertz range using diodes. While these early efforts showed promise, progress stalled for decades. Recently, however, advances in nanotechnology have reignited this area of research. Researchers discovered that tiny structures like nanometer-length-scale needle tips and plasmonic antennas could function similarly to those early diodes but at much higher frequencies.

A recent open-access study published in Science Advances by Matthew Yeung, Lu-Ting Chou, Marco Turchetti, Felix Ritzkowsky, Karl K. Berggren, and Phillip D. Keathley at MIT has demonstrated a significant step forward. They developed an electronic frequency mixer for signal detection that operates beyond 0.350 PHz using tiny nanoantennae. These nanoantennae can mix different frequencies of light, enabling analysis of signals oscillating orders of magnitude faster than the fastest accessible to conventional electronics. Such petahertz electronic devices could enable developments that ultimately revolutionize fields that require precise analysis of extremely fast optical signals, such as spectroscopy and imaging, where capturing femtosecond-scale dynamics is crucial (a femtosecond is one-millionth of one-billionth of a second).

The team’s study highlights the use of nanoantenna networks to create a broadband, on-chip electronic optical frequency mixer. This innovative approach allows for the accurate readout of optical wave forms spanning more than one octave of bandwidth. Importantly, this process worked using a commercial turnkey laser that can be purchased off the shelf, rather than a highly customized laser.

While optical frequency mixing is possible using nonlinear materials, the process is purely optical (that is, it converts light input to light output at a new frequency). Furthermore, the materials have to be many wavelengths in thickness, limiting the device size to the micrometer scale (a micrometer is one-millionth of a meter).  In contrast, the lightwave-electronic method demonstrated by the authors uses a light-driven tunneling mechanism that offers high nonlinearities for frequency mixing and direct electronic output using nanometer-scale devices (a nanometer is one-billionth of a meter).

While this study focused on characterizing light pulses of different frequencies, the researchers envision that similar devices will enable one to construct circuits using light waves. This device, with bandwidths spanning multiple octaves, could provide new ways to investigate ultrafast light-matter interactions, accelerating advancements in ultrafast source technologies. 

This work not only pushes the boundaries of what is possible in optical signal processing but also bridges the gap between the fields of electronics and optics. By connecting these two important areas of research, this study paves the way for new technologies and applications in fields like spectroscopy, imaging, and communications, ultimately advancing our ability to explore and manipulate the ultrafast dynamics of light.

The research was initially supported by the U.S. Air Force Office of Scientific Research. Ongoing research into harmonic mixing is based upon work supported by the U.S. Department of Energy, Office of Science, Office of Basic Energy Sciences. Matthew Yeung acknowledges fellowship support from MathWorks, the U.S. National Science Foundation Graduate Research Fellowship Program, and MPS-Ascend Postdoctoral Research Fellowship. Lu-Ting Chou acknowledges financial support from the China's Ministry of Education for the Overseas Internship Program from the Chinese National Science and Technology Council for the doctoral fellowship program. This work was carried out, in part, through the use of MIT.nano.

Exoplanets form in protoplanetary disks, a collection of space dust and gas orbiting a star. The leading theory of planetary formation, called core accretion, occurs when grains of dust in the disk collect and grow to form a planetary core, like a snowball rolling downhill. Once it has a strong enough gravitational pull, other material collapses around it to form the atmosphere.

A secondary theory of planetary formation is gravitational collapse. In this scenario, the disk itself becomes gravitationally unstable and collapses to form the planet, like snow being plowed into a pile. This process requires the disk to be massive, and until recently there were no known viable candidates to observe; previous research had detected the snow pile, but not what made it.

But in a new paper published today in Nature, MIT Kerr-McGee Career Development Professor Richard Teague and his colleagues report evidence that the movement of the gas surrounding the star AB Aurigae behaves as one would expect in a gravitationally unstable disk, matching numerical predictions. Their finding is akin to detecting the snowplow that made the pile. This indicates that gravitational collapse is a viable method of planetary formation. Here, Teague, who studies the formation of planetary systems in MIT’s Department of Earth, Atmospheric and Planetary Sciences (EAPS), answers a few questions about the new work.

Q: What made the AB Aurigae system a good candidate for observation?

A: There have been plenty of observations that have suggested some interesting dynamics going on the system. Groups have seen spiral arms within the disk; people have found hot spots, which some groups have interpreted as a planet; others have explained as some other instability. But it was really a disk that we knew there was lots of interesting motions going on. The data that we had previously was enough to see that it was interesting, but not really good enough to detail what was going on.

Q: What is gravitational instability when it comes to protoplanetary disks?

A: Gravitational instabilities are where the gravity from the disk itself is strong enough to perturb motions within the disk. Usually, we assume that the gravitational potential is dominated by the central star, which is the case when the mass of the disk is less than 10 percent of the stellar mass (which is most of the time). When the disk mass gets too large, gravitational potential will affect it in different ways and drive these very large spiral arms in the disk. These can have lots of different effects: They can trap the gas, they can heat it up, they can allow for angular momentum to be transported very rapidly within the disk. If it’s unstable, the disk can fragment and collapse directly to form a planet in an incredibly short period of time. Rather than the tens of thousands of years that it would take for a core accretion to happen, this would happen at a fraction of that time.

Q: How does this discovery challenge conventional wisdom around planetary formation?

A: It shows that this alternative path of forming planets via direct collapse is a way that we can form planets. This is particularly important because we’re finding more and more evidence of very large planets — say, Jupiter mass or larger — that are sitting very far away from their star. Those sorts of planets are incredibly hard to form with core accretion, because you typically need them close to the star where things happen quickly. So to form something so massive, so far away from the star is a real challenge. If we're able to show that there are sources that are massive enough that they're gravitationally unstable, this solves that problem. It's a way that perhaps newer systems can be formed, because they've always been a bit of a challenge to understand how they came about with core accretion.

Collagen, a protein found in bones and connective tissue, has been found in dinosaur fossils as old as 195 million years. That far exceeds the normal half-life of the peptide bonds that hold proteins together, which is about 500 years.

A new study from MIT offers an explanation for how collagen can survive for so much longer than expected. The research team found that a special atomic-level interaction defends collagen from attack by water molecules. This barricade prevents water from breaking the peptide bonds through a process called hydrolysis.

“We provide evidence that that interaction prevents water from attacking the peptide bonds and cleaving them. That just flies in the face of what happens with a normal peptide bond, which has a half-life of only 500 years,” says Ron Raines, the Firmenich Professor of Chemistry at MIT.

Raines is the senior author of the new study, which appears today in ACS Central Science. MIT postdoc Jinyi Yang PhD ’24 is the lead author of the paper. MIT postdoc Volga Kojasoy and graduate student Gerard Porter are also authors of the study.

Water-resistant

Collagen is the most abundant protein in animals, and it is found in not only bones but also skin, muscles, and ligaments. It’s made from long strands of protein that intertwine to form a tough triple helix.

“Collagen is the scaffold that holds us together,” Raines says. “What makes the collagen protein so stable, and such a good choice for this scaffold, is that unlike most proteins, it’s fibrous.”

In the past decade, paleobiologists have found evidence of collagen preserved in dinosaur fossils, including an 80-million-year-old Tyrannosaurus rex fossil, and a sauropodomorph fossil that is nearly 200 million years old.

Over the past 25 years, Raines’ lab has been studying collagen and how its structure enables its function. In the new study, they revealed why the peptide bonds that hold collagen together are so resistant to being broken down by water.

Peptide bonds are formed between a carbon atom from one amino acid and a nitrogen atom of the adjacent amino acid. The carbon atom also forms a double bond with an oxygen atom, forming a molecular structure called a carbonyl group. This carbonyl oxygen has a pair of electrons that don’t form bonds with any other atoms. Those electrons, the researchers found, can be shared with the carbonyl group of a neighboring peptide bond.

Because this pair of electrons is being inserted into those peptide bonds, water molecules can’t also get into the structure to disrupt the bond.

To demonstrate this, Raines and his colleagues created two interconverting mimics of collagen — the one that usually forms a triple helix, which is known as trans, and another in which the angles of the peptide bonds are rotated into a different form, known as cis. They found that the trans form of collagen did not allow water to attack and hydrolyze the bond. In the cis form, water got in and the bonds were broken.

“A peptide bond is either cis or trans, and we can change the cis to trans ratio. By doing that, we can mimic the natural state of collagen or create an unprotected peptide bond. And we saw that when it was unprotected, it was not long for the world,” Raines says.

“This work builds on a long-term effort in the Raines Group to classify the role of a long-overlooked fundamental interaction in protein structure,” says Paramjit Arora, a professor of chemistry at New York University, who was not involved in the research. “The paper directly addresses the remarkable finding of intact collagen in the ribs of a 195-million-old dinosaur fossil, and shows that overlap of filled and empty orbitals controls the conformational and hydrolytic stability of collagen.”

“No weak link”

This sharing of electrons has also been seen in protein structures known as alpha helices, which are found in many proteins. These helices may also be protected from water, but the helices are always connected by protein sequences that are more exposed, which are still susceptible to hydrolysis.

“Collagen is all triple helices, from one end to the other,” Raines says. “There’s no weak link, and that’s why I think it has survived.”

Previously, some scientists have suggested other explanations for why collagen might be preserved for millions of years, including the possibility that the bones were so dehydrated that no water could reach the peptide bonds.

“I can’t discount the contributions from other factors, but 200 million years is a long time, and I think you need something at the molecular level, at the atomic level in order to explain it,” Raines says.

The research was funded by the National Institutes of Health and the National Science Foundation.

Charging stations for electric vehicles are essential for cleaning up the transportation sector. A new study by MIT researchers suggests they’re good for business, too.

The study found that, in California, opening a charging station boosted annual spending at each nearby business by an average of about $1,500 in 2019 and about $400 between January 2021 and June 2023. The spending bump amounts to thousands of extra dollars annually for nearby businesses, with the increase particularly pronounced for businesses in underresourced areas.

The study’s authors hope the research paints a more holistic picture of the benefits of EV charging stations, beyond environmental factors.

“These increases are equal to a significant chunk of the cost of installing an EV charger, and I hope this study sheds light on these economic benefits,” says lead author Yunhan Zheng MCP ’21, SM ’21, PhD ’24, a postdoc at the Singapore-MIT Alliance for Research and Technology (SMART). “The findings could also diversify the income stream for charger providers and site hosts, and lead to more informed business models for EV charging stations.”

Zheng’s co-authors on the paper, which was published today in Nature Communications, are David Keith, a senior lecturer at the MIT Sloan School of Management; Jinhua Zhao, an MIT professor of cities and transportation; and alumni Shenhao Wang MCP ’17, SM ’17, PhD ’20 and Mi Diao MCP ’06, PhD ’10.

Understanding the EV effect

Increasing the number of electric vehicle charging stations is seen as a key prerequisite for the transition to a cleaner, electrified transportation sector. As such, the 2021 U.S. Infrastructure Investment and Jobs Act committed $7.5 billion to build a national network of public electric vehicle chargers across the U.S.

But a large amount of private investment will also be needed to make charging stations ubiquitous.

“The U.S. is investing a lot in EV chargers and really encouraging EV adoption, but many EV charging providers can’t make enough money at this stage, and getting to profitability is a major challenge,” Zheng says.

EV advocates have long argued that the presence of charging stations brings economic benefits to surrounding communities, but Zheng says previous studies on their impact relied on surveys or were small-scale. Her team of collaborators wanted to make advocates’ claims more empirical.

For their study, the researchers collected data from over 4,000 charging stations in California and 140,000 businesses, relying on anonymized credit and debit card transactions to measure changes in consumer spending. The researchers used data from 2019 through June of 2023, skipping the year 2020 to minimize the impact of the pandemic.

To judge whether charging stations caused customer spending increases, the researchers compared data from businesses within 500 meters of new charging stations before and after their installation. They also analyzed transactions from similar businesses in the same time frame that weren’t near charging stations.

Supercharging nearby businesses

The researchers found that installing a charging station boosted annual spending at nearby establishments by an average of 1.4 percent in 2019 and 0.8 percent from January 2021 to June 2023.

While that might sound like a small amount per business, it amounts to thousands of dollars in overall consumer spending increases. Specifically, those percentages translate to almost $23,000 in cumulative spending increases in 2019 and about $3,400 per year from 2021 through June 2023.

Zheng says the decline in spending increases over the two time periods might be due to a saturation of EV chargers, leading to lower utilization, as well as an overall decrease in spending per business after the Covid-19 pandemic and a reduced number of businesses served by each EV charging station in the second period. Despite this decline, the annual impact of a charging station on all its surrounding businesses would still cover approximately 11.2 percent of the average infrastructure and installation cost of a standard charging station.

Through both time frames, the spending increases were highest for businesses within about a football field’s distance from the new stations. They were also significant for businesses in disadvantaged and low-income areas, as designated by California and the Justice40 Initiative.

“The positive impacts of EV charging stations on businesses are not constrained solely to some high-income neighborhoods,” Wang says. “It highlights the importance for policymakers to develop EV charging stations in marginalized areas, because they not only foster a cleaner environment, but also serve as a catalyst for enhancing economic vitality.”

Zheng believes the findings hold a lesson for charging station developers seeking to improve the profitability of their projects.

“The joint gas station and convenience store business model could also be adopted to EV charging stations,” Zheng says. “Traditionally, many gas stations are affiliated with retail store chains, which enables owners to both sell fuel and attract customers to diversify their revenue stream. EV charging providers could consider a similar approach to internalize the positive impact of EV charging stations.”

Zheng also says the findings could support the creation of new funding models for charging stations, such as multiple businesses sharing the costs of construction so they can all benefit from the added spending.

Those changes could accelerate the creation of charging networks, but Zheng cautions that further research is needed to understand how much the study’s findings can be extrapolated to other areas. She encourages other researchers to study the economic effects of charging stations and hopes future research includes states beyond California and even other countries.

“A huge number of studies have focused on retail sales effects from traditional transportation infrastructure, such as rail and subway stations, bus stops, and street configurations,” Zhao says. “This research provides evidence for an important, emerging piece of transportation infrastructure and shows a consistently positive effect on local businesses, paving the way for future research in this area.”

The research was supported, in part, by the Singapore-MIT Alliance for Research and Technology (SMART) and the Singapore National Research Foundation. Diao was partially supported by the Natural Science Foundation of Shanghai and the Fundamental Research Funds for the Central Universities of China.

In order to train more powerful large language models, researchers use vast dataset collections that blend diverse data from thousands of web sources.

But as these datasets are combined and recombined into multiple collections, important information about their origins and restrictions on how they can be used are often lost or confounded in the shuffle.

Not only does this raise legal and ethical concerns, it can also damage a model’s performance. For instance, if a dataset is miscategorized, someone training a machine-learning model for a certain task may end up unwittingly using data that are not designed for that task.

In addition, data from unknown sources could contain biases that cause a model to make unfair predictions when deployed.

To improve data transparency, a team of multidisciplinary researchers from MIT and elsewhere launched a systematic audit of more than 1,800 text datasets on popular hosting sites. They found that more than 70 percent of these datasets omitted some licensing information, while about 50 percent had information that contained errors.

Building off these insights, they developed a user-friendly tool called the Data Provenance Explorer that automatically generates easy-to-read summaries of a dataset’s creators, sources, licenses, and allowable uses.

“These types of tools can help regulators and practitioners make informed decisions about AI deployment, and further the responsible development of AI,” says Alex “Sandy” Pentland, an MIT professor, leader of the Human Dynamics Group in the MIT Media Lab, and co-author of a new open-access paper about the project.

The Data Provenance Explorer could help AI practitioners build more effective models by enabling them to select training datasets that fit their model’s intended purpose. In the long run, this could improve the accuracy of AI models in real-world situations, such as those used to evaluate loan applications or respond to customer queries.

“One of the best ways to understand the capabilities and limitations of an AI model is understanding what data it was trained on. When you have misattribution and confusion about where data came from, you have a serious transparency issue,” says Robert Mahari, a graduate student in the MIT Human Dynamics Group, a JD candidate at Harvard Law School, and co-lead author on the paper.

Mahari and Pentland are joined on the paper by co-lead author Shayne Longpre, a graduate student in the Media Lab; Sara Hooker, who leads the research lab Cohere for AI; as well as others at MIT, the University of California at Irvine, the University of Lille in France, the University of Colorado at Boulder, Olin College, Carnegie Mellon University, Contextual AI, ML Commons, and Tidelift. The research is published today in Nature Machine Intelligence.

Focus on finetuning

Researchers often use a technique called fine-tuning to improve the capabilities of a large language model that will be deployed for a specific task, like question-answering. For finetuning, they carefully build curated datasets designed to boost a model’s performance for this one task.

The MIT researchers focused on these fine-tuning datasets, which are often developed by researchers, academic organizations, or companies and licensed for specific uses.

When crowdsourced platforms aggregate such datasets into larger collections for practitioners to use for fine-tuning, some of that original license information is often left behind.

“These licenses ought to matter, and they should be enforceable,” Mahari says.

For instance, if the licensing terms of a dataset are wrong or missing, someone could spend a great deal of money and time developing a model they might be forced to take down later because some training data contained private information.

“People can end up training models where they don’t even understand the capabilities, concerns, or risk of those models, which ultimately stem from the data,” Longpre adds.

To begin this study, the researchers formally defined data provenance as the combination of a dataset’s sourcing, creating, and licensing heritage, as well as its characteristics. From there, they developed a structured auditing procedure to trace the data provenance of more than 1,800 text dataset collections from popular online repositories.

After finding that more than 70 percent of these datasets contained “unspecified” licenses that omitted much information, the researchers worked backward to fill in the blanks. Through their efforts, they reduced the number of datasets with “unspecified” licenses to around 30 percent.

Their work also revealed that the correct licenses were often more restrictive than those assigned by the repositories.   

In addition, they found that nearly all dataset creators were concentrated in the global north, which could limit a model’s capabilities if it is trained for deployment in a different region. For instance, a Turkish language dataset created predominantly by people in the U.S. and China might not contain any culturally significant aspects, Mahari explains.

“We almost delude ourselves into thinking the datasets are more diverse than they actually are,” he says.

Interestingly, the researchers also saw a dramatic spike in restrictions placed on datasets created in 2023 and 2024, which might be driven by concerns from academics that their datasets could be used for unintended commercial purposes.

A user-friendly tool

To help others obtain this information without the need for a manual audit, the researchers built the Data Provenance Explorer. In addition to sorting and filtering datasets based on certain criteria, the tool allows users to download a data provenance card that provides a succinct, structured overview of dataset characteristics.

“We are hoping this is a step, not just to understand the landscape, but also help people going forward to make more informed choices about what data they are training on,” Mahari says.

In the future, the researchers want to expand their analysis to investigate data provenance for multimodal data, including video and speech. They also want to study how terms of service on websites that serve as data sources are echoed in datasets.

As they expand their research, they are also reaching out to regulators to discuss their findings and the unique copyright implications of fine-tuning data.

“We need data provenance and transparency from the outset, when people are creating and releasing these datasets, to make it easier for others to derive these insights,” Longpre says.

“Many proposed policy interventions assume that we can correctly assign and identify licenses associated with data, and this work first shows that this is not the case, and then significantly improves the provenance information available,” says Stella Biderman, executive director of EleutherAI, who was not involved with this work. “In addition, section 3 contains relevant legal discussion. This is very valuable to machine learning practitioners outside companies large enough to have dedicated legal teams. Many people who want to build AI systems for public good are currently quietly struggling to figure out how to handle data licensing, because the internet is not designed in a way that makes data provenance easy to figure out.”

Computer graphics and geometry processing research provide the tools needed to simulate physical phenomena like fire and flames, aiding the creation of visual effects in video games and movies as well as the fabrication of complex geometric shapes using tools like 3D printing.

Under the hood, mathematical problems called partial differential equations (PDEs) model these natural processes. Among the many PDEs used in physics and computer graphics, a class called second-order parabolic PDEs explain how phenomena can become smooth over time. The most famous example in this class is the heat equation, which predicts how heat diffuses along a surface or in a volume over time.

Researchers in geometry processing have designed numerous algorithms to solve these problems on curved surfaces, but their methods often apply only to linear problems or to a single PDE. A more general approach by researchers from MIT’s Computer Science and Artificial Intelligence Laboratory (CSAIL) tackles a general class of these potentially nonlinear problems. 

In a paper recently published in the Transactions on Graphics journal and presented at the SIGGRAPH conference, they describe an algorithm that solves different nonlinear parabolic PDEs on triangle meshes by splitting them into three simpler equations that can be solved with techniques graphics researchers already have in their software toolkit. This framework can help better analyze shapes and model complex dynamical processes.

“We provide a recipe: If you want to numerically solve a second-order parabolic PDE, you can follow a set of three steps,” says lead author Leticia Mattos Da Silva SM ’23, an MIT PhD student in electrical engineering and computer science (EECS) and CSAIL affiliate. “For each of the steps in this approach, you’re solving a simpler problem using simpler tools from geometry processing, but at the end, you get a solution to the more challenging second-order parabolic PDE.”

To accomplish this, Da Silva and her coauthors used Strang splitting, a technique that allows geometry processing researchers to break the PDE down into problems they know how to solve efficiently.

First, their algorithm advances a solution forward in time by solving the heat equation (also called the “diffusion equation”), which models how heat from a source spreads over a shape. Picture using a blow torch to warm up a metal plate — this equation describes how heat from that spot would diffuse over it. 
This step can be completed easily with linear algebra.

Now, imagine that the parabolic PDE has additional nonlinear behaviors that are not described by the spread of heat. This is where the second step of the algorithm comes in: it accounts for the nonlinear piece by solving a Hamilton-Jacobi (HJ) equation, a first-order nonlinear PDE. 

While generic HJ equations can be hard to solve, Mattos Da Silva and coauthors prove that their splitting method applied to many important PDEs yields an HJ equation that can be solved via convex optimization algorithms. Convex optimization is a standard tool for which researchers in geometry processing already have efficient and reliable software. In the final step, the algorithm advances a solution forward in time using the heat equation again to advance the more complex second-order parabolic PDE forward in time.


Among other applications, the framework could help simulate fire and flames more efficiently. “There’s a huge pipeline that creates a video with flames being simulated, but at the heart of it is a PDE solver,” says Mattos Da Silva. For these pipelines, an essential step is solving the G-equation, a nonlinear parabolic PDE that models the front propagation of the flame and can be solved using the researchers’ framework.

The team’s algorithm can also solve the diffusion equation in the logarithmic domain, where it becomes nonlinear. Senior author Justin Solomon, associate professor of EECS and leader of the CSAIL Geometric Data Processing Group, previously developed a state-of-the-art technique for optimal transport that requires taking the logarithm of the result of heat diffusion. Mattos Da Silva’s framework provided more reliable computations by doing diffusion directly in the logarithmic domain. This enabled a more stable way to, for example, find a geometric notion of average among distributions on surface meshes like a model of a koala.

Even though their framework focuses on general, nonlinear problems, it can also be used to solve linear PDE. For instance, the method solves the Fokker-Planck equation, where heat diffuses in a linear way, but there are additional terms that drift in the same direction heat is spreading. In a straightforward application, the approach modeled how swirls would evolve over the surface of a triangulated sphere. The result resembles purple-and-brown latte art.

The researchers note that this project is a starting point for tackling the nonlinearity in other PDEs that appear in graphics and geometry processing head-on. For example, they focused on static surfaces but would like to apply their work to moving ones, too. Moreover, their framework solves problems involving a single parabolic PDE, but the team would also like to tackle problems involving coupled parabolic PDE. These types of problems arise in biology and chemistry, where the equation describing the evolution of each agent in a mixture, for example, is linked to the others’ equations.

Mattos Da Silva and Solomon wrote the paper with Oded Stein, assistant professor at the University of Southern California’s Viterbi School of Engineering. Their work was supported, in part, by an MIT Schwarzman College of Computing Fellowship funded by Google, a MathWorks Fellowship, the Swiss National Science Foundation, the U.S. Army Research Office, the U.S. Air Force Office of Scientific Research, the U.S. National Science Foundation, MIT-IBM Watson AI Lab, the Toyota-CSAIL Joint Research Center, Adobe Systems, and Google Research.

Using functional magnetic resonance imaging (fMRI), neuroscientists have identified several regions of the brain that are responsible for processing language. However, discovering the specific functions of neurons in those regions has proven difficult because fMRI, which measures changes in blood flow, doesn’t have high enough resolution to reveal what small populations of neurons are doing.

Now, using a more precise technique that involves recording electrical activity directly from the brain, MIT neuroscientists have identified different clusters of neurons that appear to process different amounts of linguistic context. These “temporal windows” range from just one word up to about six words.

The temporal windows may reflect different functions for each population, the researchers say. Populations with shorter windows may analyze the meanings of individual words, while those with longer windows may interpret more complex meanings created when words are strung together.

“This is the first time we see clear heterogeneity within the language network,” says Evelina Fedorenko, an associate professor of neuroscience at MIT. “Across dozens of fMRI experiments, these brain areas all seem to do the same thing, but it’s a large, distributed network, so there’s got to be some structure there. This is the first clear demonstration that there is structure, but the different neural populations are spatially interleaved so we can’t see these distinctions with fMRI.”

Fedorenko, who is also a member of MIT’s McGovern Institute for Brain Research, is the senior author of the study, which appears today in Nature Human Behavior. MIT postdoc Tamar Regev and Harvard University graduate student Colton Casto are the lead authors of the paper.

Temporal windows

Functional MRI, which has helped scientists learn a great deal about the roles of different parts of the brain, works by measuring changes in blood flow in the brain. These measurements act as a proxy of neural activity during a particular task. However, each “voxel,” or three-dimensional chunk, of an fMRI image represents hundreds of thousands to millions of neurons and sums up activity across about two seconds, so it can’t reveal fine-grained detail about what those neurons are doing.

One way to get more detailed information about neural function is to record electrical activity using electrodes implanted in the brain. These data are hard to come by because this procedure is done only in patients who are already undergoing surgery for a neurological condition such as severe epilepsy.

“It can take a few years to get enough data for a task because these patients are relatively rare, and in a given patient electrodes are implanted in idiosyncratic locations based on clinical needs, so it takes a while to assemble a dataset with sufficient coverage of some target part of the cortex. But these data, of course, are the best kind of data we can get from human brains: You know exactly where you are spatially and you have very fine-grained temporal information,” Fedorenko says.

In a 2016 study, Fedorenko reported using this approach to study the language processing regions of six people. Electrical activity was recorded while the participants read four different types of language stimuli: complete sentences, lists of words, lists of non-words, and “jabberwocky” sentences — sentences that have grammatical structure but are made of nonsense words.

Those data showed that in some neural populations in language processing regions, activity would gradually build up over a period of several words, when the participants were reading sentences. However, this did not happen when they read lists of words, lists of nonwords, of Jabberwocky sentences.

In the new study, Regev and Casto went back to those data and analyzed the temporal response profiles in greater detail. In their original dataset, they had recordings of electrical activity from 177 language-responsive electrodes across the six patients. Conservative estimates suggest that each electrode represents an average of activity from about 200,000 neurons. They also obtained new data from a second set of 16 patients, which included recordings from another 362 language-responsive electrodes.

When the researchers analyzed these data, they found that in some of the neural populations, activity would fluctuate up and down with each word. In others, however, activity would build up over multiple words before falling again, and yet others would show a steady buildup of neural activity over longer spans of words.

By comparing their data with predictions made by a computational model that the researchers designed to process stimuli with different temporal windows, the researchers found that neural populations from language processing areas could be divided into three clusters. These clusters represent temporal windows of either one, four, or six words.

“It really looks like these neural populations integrate information across different timescales along the sentence,” Regev says.

Processing words and meaning

These differences in temporal window size would have been impossible to see using fMRI, the researchers say.

“At the resolution of fMRI, we don’t see much heterogeneity within language-responsive regions. If you localize in individual participants the voxels in their brain that are most responsive to language, you find that their responses to sentences, word lists, jabberwocky sentences and non-word lists are highly similar,” Casto says.

The researchers were also able to determine the anatomical locations where these clusters were found. Neural populations with the shortest temporal window were found predominantly in the posterior temporal lobe, though some were also found in the frontal or anterior temporal lobes. Neural populations from the two other clusters, with longer temporal windows, were spread more evenly throughout the temporal and frontal lobes.

Fedorenko’s lab now plans to study whether these timescales correspond to different functions. One possibility is that the shortest timescale populations may be processing the meanings of a single word, while those with longer timescales interpret the meanings represented by multiple words.

“We already know that in the language network, there is sensitivity to how words go together and to the meanings of individual words,” Regev says. “So that could potentially map to what we’re finding, where the longest timescale is sensitive to things like syntax or relationships between words, and maybe the shortest timescale is more sensitive to features of single words or parts of them.”

The research was funded by the Zuckerman-CHE STEM Leadership Program, the Poitras Center for Psychiatric Disorders Research, the Kempner Institute for the Study of Natural and Artificial Intelligence at Harvard University, the U.S. National Institutes of Health, an American Epilepsy Society Research and Training Fellowship, the McDonnell Center for Systems Neuroscience, Fondazione Neurone, the McGovern Institute, MIT’s Department of Brain and Cognitive Sciences, and the Simons Center for the Social Brain.

For the past decade, disordered rock salt has been studied as a potential breakthrough cathode material for use in lithium-ion batteries and a key to creating low-cost, high-energy storage for everything from cell phones to electric vehicles to renewable energy storage.

A new MIT study is making sure the material fulfills that promise.

Led by Ju Li, the Tokyo Electric Power Company Professor in Nuclear Engineering and professor of materials science and engineering, a team of researchers describe a new class of partially disordered rock salt cathode, integrated with polyanions — dubbed disordered rock salt-polyanionic spinel, or DRXPS — that delivers high energy density at high voltages with significantly improved cycling stability.

“There is typically a trade-off in cathode materials between energy density and cycling stability … and with this work we aim to push the envelope by designing new cathode chemistries,” says Yimeng Huang, a postdoc in the Department of Nuclear Science and Engineering and first author of a paper describing the work published today in Nature Energy. “(This) material family has high energy density and good cycling stability because it integrates two major types of cathode materials, rock salt and polyanionic olivine, so it has the benefits of both.”

Importantly, Li adds, the new material family is primarily composed of manganese, an earth-abundant element that is significantly less expensive than elements like nickel and cobalt, which are typically used in cathodes today.

“Manganese is at least five times less expensive than nickel, and about 30 times less expensive than cobalt,” Li says. “Manganese is also the one of the keys to achieving higher energy densities, so having that material be much more earth-abundant is a tremendous advantage.”

A possible path to renewable energy infrastructure

That advantage will be particularly critical, Li and his co-authors wrote, as the world looks to build the renewable energy infrastructure needed for a low- or no-carbon future.

Batteries are a particularly important part of that picture, not only for their potential to decarbonize transportation with electric cars, buses, and trucks, but also because they will be essential to addressing the intermittency issues of wind and solar power by storing excess energy, then feeding it back into the grid at night or on calm days, when renewable generation drops.

Given the high cost and relative rarity of materials like cobalt and nickel, they wrote, efforts to rapidly scale up electric storage capacity would likely lead to extreme cost spikes and potentially significant materials shortages.

“If we want to have true electrification of energy generation, transportation, and more, we need earth-abundant batteries to store intermittent photovoltaic and wind power,” Li says. “I think this is one of the steps toward that dream.”

That sentiment was shared by Gerbrand Ceder, the Samsung Distinguished Chair in Nanoscience and Nanotechnology Research and a professor of materials science and engineering at the University of California at Berkeley.

“Lithium-ion batteries are a critical part of the clean energy transition,” Ceder says. “Their continued growth and price decrease depends on the development of inexpensive, high-performance cathode materials made from earth-abundant materials, as presented in this work.”

Overcoming obstacles in existing materials

The new study addresses one of the major challenges facing disordered rock salt cathodes — oxygen mobility.

While the materials have long been recognized for offering very high capacity — as much as 350 milliampere-hour per gram — as compared to traditional cathode materials, which typically have capacities of between 190 and 200 milliampere-hour per gram, it is not very stable.

The high capacity is contributed partially by oxygen redox, which is activated when the cathode is charged to high voltages. But when that happens, oxygen becomes mobile, leading to reactions with the electrolyte and degradation of the material, eventually leaving it effectively useless after prolonged cycling.

To overcome those challenges, Huang added another element — phosphorus — that essentially acts like a glue, holding the oxygen in place to mitigate degradation.

“The main innovation here, and the theory behind the design, is that Yimeng added just the right amount of phosphorus, formed so-called polyanions with its neighboring oxygen atoms, into a cation-deficient rock salt structure that can pin them down,” Li explains. “That allows us to basically stop the percolating oxygen transport due to strong covalent bonding between phosphorus and oxygen … meaning we can both utilize the oxygen-contributed capacity, but also have good stability as well.”

That ability to charge batteries to higher voltages, Li says, is crucial because it allows for simpler systems to manage the energy they store.

“You can say the quality of the energy is higher,” he says. “The higher the voltage per cell, then the less you need to connect them in series in the battery pack, and the simpler the battery management system.”

Pointing the way to future studies

While the cathode material described in the study could have a transformative impact on lithium-ion battery technology, there are still several avenues for study going forward.

Among the areas for future study, Huang says, are efforts to explore new ways to fabricate the material, particularly for morphology and scalability considerations.

“Right now, we are using high-energy ball milling for mechanochemical synthesis, and … the resulting morphology is non-uniform and has small average particle size (about 150 nanometers). This method is also not quite scalable,” he says. “We are trying to achieve a more uniform morphology with larger particle sizes using some alternate synthesis methods, which would allow us to increase the volumetric energy density of the material and may allow us to explore some coating methods … which could further improve the battery performance. The future methods, of course, should be industrially scalable.”

In addition, he says, the disordered rock salt material by itself is not a particularly good conductor, so significant amounts of carbon — as much as 20 weight percent of the cathode paste — were added to boost its conductivity. If the team can reduce the carbon content in the electrode without sacrificing performance, there will be higher active material content in a battery, leading to an increased practical energy density.

“In this paper, we just used Super P, a typical conductive carbon consisting of nanospheres, but they’re not very efficient,” Huang says. “We are now exploring using carbon nanotubes, which could reduce the carbon content to just 1 or 2 weight percent, which could allow us to dramatically increase the amount of the active cathode material.”

Aside from decreasing carbon content, making thick electrodes, he adds, is yet another way to increase the practical energy density of the battery. This is another area of research that the team is working on.

“This is only the beginning of DRXPS research, since we only explored a few chemistries within its vast compositional space,” he continues. “We can play around with different ratios of lithium, manganese, phosphorus, and oxygen, and with various combinations of other polyanion-forming elements such as boron, silicon, and sulfur.”

With optimized compositions, more scalable synthesis methods, better morphology that allows for uniform coatings, lower carbon content, and thicker electrodes, he says, the DRXPS cathode family is very promising in applications of electric vehicles and grid storage, and possibly even in consumer electronics, where the volumetric energy density is very important.

This work was supported with funding from the Honda Research Institute USA Inc. and the Molecular Foundry at Lawrence Berkeley National Laboratory, and used resources of the National Synchrotron Light Source II at Brookhaven National Laboratory and the Advanced Photon Source at Argonne National Laboratory. The work was carried out, in part, using MIT.nano’s facilities. 

The most recent email you sent was likely encrypted using a tried-and-true method that relies on the idea that even the fastest computer would be unable to efficiently break a gigantic number into factors.

Quantum computers, on the other hand, promise to rapidly crack complex cryptographic systems that a classical computer might never be able to unravel. This promise is based on a quantum factoring algorithm proposed in 1994 by Peter Shor, who is now a professor at MIT.

But while researchers have taken great strides in the last 30 years, scientists have yet to build a quantum computer powerful enough to run Shor’s algorithm.

As some researchers work to build larger quantum computers, others have been trying to improve Shor’s algorithm so it could run on a smaller quantum circuit. About a year ago, New York University computer scientist Oded Regev proposed a major theoretical improvement. His algorithm could run faster, but the circuit would require more memory.

Building off those results, MIT researchers have proposed a best-of-both-worlds approach that combines the speed of Regev’s algorithm with the memory-efficiency of Shor’s. This new algorithm is as fast as Regev’s, requires fewer quantum building blocks known as qubits, and has a higher tolerance to quantum noise, which could make it more feasible to implement in practice.

In the long run, this new algorithm could inform the development of novel encryption methods that can withstand the code-breaking power of quantum computers.

“If large-scale quantum computers ever get built, then factoring is toast and we have to find something else to use for cryptography. But how real is this threat? Can we make quantum factoring practical? Our work could potentially bring us one step closer to a practical implementation,” says Vinod Vaikuntanathan, the Ford Foundation Professor of Engineering, a member of the Computer Science and Artificial Intelligence Laboratory (CSAIL), and senior author of a paper describing the algorithm.

The paper’s lead author is Seyoon Ragavan, a graduate student in the MIT Department of Electrical Engineering and Computer Science. The research will be presented at the 2024 International Cryptology Conference.

Cracking cryptography

To securely transmit messages over the internet, service providers like email clients and messaging apps typically rely on RSA, an encryption scheme invented by MIT researchers Ron Rivest, Adi Shamir, and Leonard Adleman in the 1970s (hence the name “RSA”). The system is based on the idea that factoring a 2,048-bit integer (a number with 617 digits) is too hard for a computer to do in a reasonable amount of time.

That idea was flipped on its head in 1994 when Shor, then working at Bell Labs, introduced an algorithm which proved that a quantum computer could factor quickly enough to break RSA cryptography.

“That was a turning point. But in 1994, nobody knew how to build a large enough quantum computer. And we’re still pretty far from there. Some people wonder if they will ever be built,” says Vaikuntanathan.

It is estimated that a quantum computer would need about 20 million qubits to run Shor’s algorithm. Right now, the largest quantum computers have around 1,100 qubits.

A quantum computer performs computations using quantum circuits, just like a classical computer uses classical circuits. Each quantum circuit is composed of a series of operations known as quantum gates. These quantum gates utilize qubits, which are the smallest building blocks of a quantum computer, to perform calculations.

But quantum gates introduce noise, so having fewer gates would improve a machine’s performance. Researchers have been striving to enhance Shor’s algorithm so it could be run on a smaller circuit with fewer quantum gates.

That is precisely what Regev did with the circuit he proposed a year ago.

“That was big news because it was the first real improvement to Shor’s circuit from 1994,” Vaikuntanathan says.

The quantum circuit Shor proposed has a size proportional to the square of the number being factored. That means if one were to factor a 2,048-bit integer, the circuit would need millions of gates.

Regev’s circuit requires significantly fewer quantum gates, but it needs many more qubits to provide enough memory. This presents a new problem.

“In a sense, some types of qubits are like apples or oranges. If you keep them around, they decay over time. You want to minimize the number of qubits you need to keep around,” explains Vaikuntanathan.

He heard Regev speak about his results at a workshop last August. At the end of his talk, Regev posed a question: Could someone improve his circuit so it needs fewer qubits? Vaikuntanathan and Ragavan took up that question.

Quantum ping-pong

To factor a very large number, a quantum circuit would need to run many times, performing operations that involve computing powers, like 2 to the power of 100.

But computing such large powers is costly and difficult to perform on a quantum computer, since quantum computers can only perform reversible operations. Squaring a number is not a reversible operation, so each time a number is squared, more quantum memory must be added to compute the next square.

The MIT researchers found a clever way to compute exponents using a series of Fibonacci numbers that requires simple multiplication, which is reversible, rather than squaring. Their method needs just two quantum memory units to compute any exponent.

“It is kind of like a ping-pong game, where we start with a number and then bounce back and forth, multiplying between two quantum memory registers,” Vaikuntanathan adds.

They also tackled the challenge of error correction. The circuits proposed by Shor and Regev require every quantum operation to be correct for their algorithm to work, Vaikuntanathan says. But error-free quantum gates would be infeasible on a real machine.

They overcame this problem using a technique to filter out corrupt results and only process the right ones.

The end-result is a circuit that is significantly more memory-efficient. Plus, their error correction technique would make the algorithm more practical to deploy.

“The authors resolve the two most important bottlenecks in the earlier quantum factoring algorithm. Although still not immediately practical, their work brings quantum factoring algorithms closer to reality,” adds Regev.

In the future, the researchers hope to make their algorithm even more efficient and, someday, use it to test factoring on a real quantum circuit.

“The elephant-in-the-room question after this work is: Does it actually bring us closer to breaking RSA cryptography? That is not clear just yet; these improvements currently only kick in when the integers are much larger than 2,048 bits. Can we push this algorithm and make it more feasible than Shor’s even for 2,048-bit integers?” says Ragavan.

This work is funded by an Akamai Presidential Fellowship, the U.S. Defense Advanced Research Projects Agency, the National Science Foundation, the MIT-IBM Watson AI Lab, a Thornton Family Faculty Research Innovation Fellowship, and a Simons Investigator Award.

Low-calorie diets and intermittent fasting have been shown to have numerous health benefits: They can delay the onset of some age-related diseases and lengthen lifespan, not only in humans but many other organisms.

Many complex mechanisms underlie this phenomenon. Previous work from MIT has shown that one way fasting exerts its beneficial effects is by boosting the regenerative abilities of intestinal stem cells, which helps the intestine recover from injuries or inflammation.

In a study of mice, MIT researchers have now identified the pathway that enables this enhanced regeneration, which is activated once the mice begin “refeeding” after the fast. They also found a downside to this regeneration: When cancerous mutations occurred during the regenerative period, the mice were more likely to develop early-stage intestinal tumors.

“Having more stem cell activity is good for regeneration, but too much of a good thing over time can have less favorable consequences,” says Omer Yilmaz, an MIT associate professor of biology, a member of MIT’s Koch Institute for Integrative Cancer Research, and the senior author of the new study.

Yilmaz adds that further studies are needed before forming any conclusion as to whether fasting has a similar effect in humans.

“We still have a lot to learn, but it is interesting that being in either the state of fasting or refeeding when exposure to mutagen occurs can have a profound impact on the likelihood of developing a cancer in these well-defined mouse models,” he says.

MIT postdocs Shinya Imada and Saleh Khawaled are the lead authors of the paper, which appears today in Nature.

Driving regeneration

For several years, Yilmaz’s lab has been investigating how fasting and low-calorie diets affect intestinal health. In a 2018 study, his team reported that during a fast, intestinal stem cells begin to use lipids as an energy source, instead of carbohydrates. They also showed that fasting led to a significant boost in stem cells’ regenerative ability.

However, unanswered questions remained: How does fasting trigger this boost in regenerative ability, and when does the regeneration begin?

“Since that paper, we’ve really been focused on understanding what is it about fasting that drives regeneration,” Yilmaz says. “Is it fasting itself that’s driving regeneration, or eating after the fast?”

In their new study, the researchers found that stem cell regeneration is suppressed during fasting but then surges during the refeeding period. The researchers followed three groups of mice — one that fasted for 24 hours, another one that fasted for 24 hours and then was allowed to eat whatever they wanted during a 24-hour refeeding period, and a control group that ate whatever they wanted throughout the experiment.

The researchers analyzed intestinal stem cells’ ability to proliferate at different time points and found that the stem cells showed the highest levels of proliferation at the end of the 24-hour refeeding period. These cells were also more proliferative than intestinal stem cells from mice that had not fasted at all.

“We think that fasting and refeeding represent two distinct states,” Imada says. “In the fasted state, the ability of cells to use lipids and fatty acids as an energy source enables them to survive when nutrients are low. And then it’s the postfast refeeding state that really drives the regeneration. When nutrients become available, these stem cells and progenitor cells activate programs that enable them to build cellular mass and repopulate the intestinal lining.”

Further studies revealed that these cells activate a cellular signaling pathway known as mTOR, which is involved in cell growth and metabolism. One of mTOR’s roles is to regulate the translation of messenger RNA into protein, so when it’s activated, cells produce more protein. This protein synthesis is essential for stem cells to proliferate.

The researchers showed that mTOR activation in these stem cells also led to production of large quantities of polyamines — small molecules that help cells to grow and divide.

“In the refed state, you’ve got more proliferation, and you need to build cellular mass. That requires more protein, to build new cells, and those stem cells go on to build more differentiated cells or specialized intestinal cell types that line the intestine,” Khawaled says.

Too much of a good thing

The researchers also found that when stem cells are in this highly regenerative state, they are more prone to become cancerous. Intestinal stem cells are among the most actively dividing cells in the body, as they help the lining of the intestine completely turn over every five to 10 days. Because they divide so frequently, these stem cells are the most common source of precancerous cells in the intestine.

In this study, the researchers discovered that if they turned on a cancer-causing gene in the mice during the refeeding stage, they were much more likely to develop precancerous polyps than if the gene was turned on during the fasting state. Cancer-linked mutations that occurred during the refeeding state were also much more likely to produce polyps than mutations that occurred in mice that did not undergo the cycle of fasting and refeeding.

“I want to emphasize that this was all done in mice, using very well-defined cancer mutations. In humans it’s going to be a much more complex state,” Yilmaz says. “But it does lead us to the following notion: Fasting is very healthy, but if you’re unlucky and you’re refeeding after a fasting, and you get exposed to a mutagen, like a charred steak or something, you might actually be increasing your chances of developing a lesion that can go on to give rise to cancer.”

Yilmaz also noted that the regenerative benefits of fasting could be significant for people who undergo radiation treatment, which can damage the intestinal lining, or other types of intestinal injury. His lab is now studying whether polyamine supplements could help to stimulate this kind of regeneration, without the need to fast.

“This fascinating study provides insights into the complex interplay between food consumption, stem cell biology, and cancer risk,” says Ophir Klein, a professor of medicine at the University of California at San Francisco and Cedars-Sinai Medical Center, who was not involved in the study. “Their work lays a foundation for testing polyamines as compounds that may augment intestinal repair after injuries, and it suggests that careful consideration is needed when planning diet-based strategies for regeneration to avoid increasing cancer risk.”

The research was funded, in part, by Pew-Stewart Scholars Program for Cancer Research award, the MIT Stem Cell Initiative, the Koch Institute Frontier Research Program via the Kathy and Curt Marble Cancer Research Fund, and the Bridge Project, a partnership between the Koch Institute for Integrative Cancer Research at MIT and the Dana-Farber/Harvard Cancer Center.

The blades of propellers and wind turbines are designed based on aerodynamics principles that were first described mathematically more than a century ago. But engineers have long realized that these formulas don’t work in every situation. To compensate, they have added ad hoc “correction factors” based on empirical observations.

Now, for the first time, engineers at MIT have developed a comprehensive, physics-based model that accurately represents the airflow around rotors even under extreme conditions, such as when the blades are operating at high forces and speeds, or are angled in certain directions. The model could improve the way rotors themselves are designed, but also the way wind farms are laid out and operated. The new findings are described today in the journal Nature Communications, in an open-access paper by MIT postdoc Jaime Liew, doctoral student Kirby Heck, and Michael Howland, the Esther and Harold E. Edgerton Assistant Professor of Civil and Environmental Engineering.

“We’ve developed a new theory for the aerodynamics of rotors,” Howland says. This theory can be used to determine the forces, flow velocities, and power of a rotor, whether that rotor is extracting energy from the airflow, as in a wind turbine, or applying energy to the flow, as in a ship or airplane propeller. “The theory works in both directions,” he says.

Because the new understanding is a fundamental mathematical model, some of its implications could potentially be applied right away. For example, operators of wind farms must constantly adjust a variety of parameters, including the orientation of each turbine as well as its rotation speed and the angle of its blades, in order to maximize power output while maintaining safety margins. The new model can provide a simple, speedy way of optimizing those factors in real time.

“This is what we’re so excited about, is that it has immediate and direct potential for impact across the value chain of wind power,” Howland says.

Modeling the momentum

Known as momentum theory, the previous model of how rotors interact with their fluid environment — air, water, or otherwise — was initially developed late in the 19th century. With this theory, engineers can start with a given rotor design and configuration, and determine the maximum amount of power that can be derived from that rotor — or, conversely, if it’s a propeller, how much power is needed to generate a given amount of propulsive force.

Momentum theory equations “are the first thing you would read about in a wind energy textbook, and are the first thing that I talk about in my classes when I teach about wind power,” Howland says. From that theory, physicist Albert Betz calculated in 1920 the maximum amount of energy that could theoretically be extracted from wind. Known as the Betz limit, this amount is 59.3 percent of the kinetic energy of the incoming wind.

But just a few years later, others found that the momentum theory broke down “in a pretty dramatic way” at higher forces that correspond to faster blade rotation speeds or different blade angles, Howland says. It fails to predict not only the amount, but even the direction of changes in thrust force at higher rotation speeds or different blade angles: Whereas the theory said the force should start going down above a certain rotation speed or blade angle, experiments show the opposite — that the force continues to increase. “So, it’s not just quantitatively wrong, it’s qualitatively wrong,” Howland says.

The theory also breaks down when there is any misalignment between the rotor and the airflow, which Howland says is “ubiquitous” on wind farms, where turbines are constantly adjusting to changes in wind directions. In fact, in an earlier paper in 2022, Howland and his team found that deliberately misaligning some turbines slightly relative to the incoming airflow within a wind farm significantly improves the overall power output of the wind farm by reducing wake disturbances to the downstream turbines.

In the past, when designing the profile of rotor blades, the layout of wind turbines in a farm, or the day-to-day operation of wind turbines, engineers have relied on ad hoc adjustments added to the original mathematical formulas, based on some wind tunnel tests and experience with operating wind farms, but with no theoretical underpinnings.

Instead, to arrive at the new model, the team analyzed the interaction of airflow and turbines using detailed computational modeling of the aerodynamics. They found that, for example, the original model had assumed that a drop in air pressure immediately behind the rotor would rapidly return to normal ambient pressure just a short way downstream. But it turns out, Howland says, that as the thrust force keeps increasing, “that assumption is increasingly inaccurate.”

And the inaccuracy occurs very close to the point of the Betz limit that theoretically predicts the maximum performance of a turbine — and therefore is just the desired operating regime for the turbines. “So, we have Betz’s prediction of where we should operate turbines, and within 10 percent of that operational set point that we think maximizes power, the theory completely deteriorates and doesn’t work,” Howland says.

Through their modeling, the researchers also found a way to compensate for the original formula’s reliance on a one-dimensional modeling that assumed the rotor was always precisely aligned with the airflow. To do so, they used fundamental equations that were developed to predict the lift of three-dimensional wings for aerospace applications.

The researchers derived their new model, which they call a unified momentum model, based on theoretical analysis, and then validated it using computational fluid dynamics modeling. In followup work not yet published, they are doing further validation using wind tunnel and field tests.

Fundamental understanding

One interesting outcome of the new formula is that it changes the calculation of the Betz limit, showing that it’s possible to extract a bit more power than the original formula predicted. Although it’s not a significant change — on the order of a few percent — “it’s interesting that now we have a new theory, and the Betz limit that’s been the rule of thumb for a hundred years is actually modified because of the new theory,” Howland says. “And that’s immediately useful.” The new model shows how to maximize power from turbines that are misaligned with the airflow, which the Betz limit cannot account for.

The aspects related to controlling both individual turbines and arrays of turbines can be implemented without requiring any modifications to existing hardware in place within wind farms. In fact, this has already happened, based on earlier work from Howland and his collaborators two years ago that dealt with the wake interactions between turbines in a wind farm, and was based on the existing, empirically based formulas.

“This breakthrough is a natural extension of our previous work on optimizing utility-scale wind farms,” he says, because in doing that analysis, they saw the shortcomings of the existing methods for analyzing the forces at work and predicting power produced by wind turbines. “Existing modeling using empiricism just wasn’t getting the job done,” he says.

In a wind farm, individual turbines will sap some of the energy available to neighboring turbines, because of wake effects. Accurate wake modeling is important both for designing the layout of turbines in a wind farm, and also for the operation of that farm, determining moment to moment how to set the angles and speeds of each turbine in the array.

Until now, Howland says, even the operators of wind farms, the manufacturers, and the designers of the turbine blades had no way to predict how much the power output of a turbine would be affected by a given change such as its angle to the wind without using empirical corrections. “That’s because there was no theory for it. So, that’s what we worked on here. Our theory can directly tell you, without any empirical corrections, for the first time, how you should actually operate a wind turbine to maximize its power,” he says.

Because the fluid flow regimes are similar, the model also applies to propellers, whether for aircraft or ships, and also for hydrokinetic turbines such as tidal or river turbines. Although they didn’t focus on that aspect in this research, “it’s in the theoretical modeling naturally,” he says.

The new theory exists in the form of a set of mathematical formulas that a user could incorporate in their own software, or as an open-source software package that can be freely downloaded from GitHub. “It’s an engineering model developed for fast-running tools for rapid prototyping and control and optimization,” Howland says. “The goal of our modeling is to position the field of wind energy research to move more aggressively in the development of the wind capacity and reliability necessary to respond to climate change.”

The work was supported by the National Science Foundation and Siemens Gamesa Renewable Energy.

Legal documents are notoriously difficult to understand, even for lawyers. This raises the question: Why are these documents written in a style that makes them so impenetrable?

MIT cognitive scientists believe they have uncovered the answer to that question. Just as “magic spells” use special rhymes and archaic terms to signal their power, the convoluted language of legalese acts to convey a sense of authority, they conclude.

In a study appearing this week in the journal of the Proceedings of the National Academy of Sciences, the researchers found that even non-lawyers use this type of language when asked to write laws.

“People seem to understand that there’s an implicit rule that this is how laws should sound, and they write them that way,” says Edward Gibson, an MIT professor of brain and cognitive sciences and the senior author of the study.

Eric Martinez PhD ’24 is the lead author of the study. Francis Mollica, a lecturer at the University of Melbourne, is also an author of the paper.

Casting a legal spell

Gibson’s research group has been studying the unique characteristics of legalese since 2020, when Martinez came to MIT after earning a law degree from Harvard Law School. In a 2022 study, Gibson, Martinez, and Mollica analyzed legal contracts totaling about 3.5 million words, comparing them with other types of writing, including movie scripts, newspaper articles, and academic papers.

That analysis revealed that legal documents frequently have long definitions inserted in the middle of sentences — a feature known as “center-embedding.” Linguists have previously found that this kind of structure can make text much more difficult to understand.

“Legalese somehow has developed this tendency to put structures inside other structures, in a way which is not typical of human languages,” Gibson says.

In a follow-up study published in 2023, the researchers found that legalese also makes documents more difficult for lawyers to understand. Lawyers tended to prefer plain English versions of documents, and they rated those versions to be just as enforceable as traditional legal documents.

“Lawyers also find legalese to be unwieldy and complicated,” Gibson says. “Lawyers don’t like it, laypeople don’t like it, so the point of this current paper was to try and figure out why they write documents this way.”

The researchers had a couple of hypotheses for why legalese is so prevalent. One was the “copy and edit hypothesis,” which suggests that legal documents begin with a simple premise, and then additional information and definitions are inserted into already existing sentences, creating complex center-embedded clauses.

“We thought it was plausible that what happens is you start with an initial draft that’s simple, and then later you think of all these other conditions that you want to include. And the idea is that once you’ve started, it’s much easier to center-embed that into the existing provision,” says Martinez, who is now a fellow and instructor at the University of Chicago Law School.

However, the findings ended up pointing toward a different hypothesis, the so-called “magic spell hypothesis.” Just as magic spells are written with a distinctive style that sets them apart from everyday language, the convoluted style of legal language appears to signal a special kind of authority, the researchers say.

“In English culture, if you want to write something that’s a magic spell, people know that the way to do that is you put a lot of old-fashioned rhymes in there. We think maybe center-embedding is signaling legalese in the same way,” Gibson says.

In this study, the researchers asked about 200 non-lawyers (native speakers of English living in the United States, who were recruited through a crowdsourcing site called Prolific), to write two types of texts. In the first task, people were told to write laws prohibiting crimes such as drunk driving, burglary, arson, and drug trafficking. In the second task, they were asked to write stories about those crimes.

To test the copy and edit hypothesis, half of the participants were asked to add additional information after they wrote their initial law or story. The researchers found that all of the subjects wrote laws with center-embedded clauses, regardless of whether they wrote the law all at once or were told to write a draft and then add to it later. And, when they wrote stories related to those laws, they wrote in much plainer English, regardless of whether they had to add information later.

“When writing laws, they did a lot of center-embedding regardless of whether or not they had to edit it or write it from scratch. And in that narrative text, they did not use center-embedding in either case,” Martinez says.

In another set of experiments, about 80 participants were asked to write laws, as well as descriptions that would explain those laws to visitors from another country. In these experiments, participants again used center-embedding for their laws, but not for the descriptions of those laws.

The origins of legalese

Gibson’s lab is now investigating the origins of center-embedding in legal documents. Early American laws were based on British law, so the researchers plan to analyze British laws to see if they feature the same kind of grammatical construction. And going back much farther, they plan to analyze whether center-embedding is found in the Hammurabi Code, the earliest known set of laws, which dates to around 1750 BC.

“There may be just a stylistic way of writing from back then, and if it was seen as successful, people would use that style in other languages,” Gibson says. “I would guess that it’s an accidental property of how the laws were written the first time, but we don’t know that yet.”

The researchers hope that their work, which has identified specific aspects of legal language that make it more difficult to understand, will motivate lawmakers to try to make laws more comprehensible. Efforts to write legal documents in plainer language date to at least the 1970s, when President Richard Nixon declared that federal regulations should be written in “layman’s terms.” However, legal language has changed very little since that time.

“We have learned only very recently what it is that makes legal language so complicated, and therefore I am optimistic about being able to change it,” Gibson says. 

For many decades, nuclear fusion power has been viewed as the ultimate energy source. A fusion power plant could generate carbon-free energy at a scale needed to address climate change. And it could be fueled by deuterium recovered from an essentially endless source — seawater.

Decades of work and billions of dollars in research funding have yielded many advances, but challenges remain. To Ju Li, the TEPCO Professor in Nuclear Science and Engineering and a professor of materials science and engineering at MIT, there are still two big challenges. The first is to build a fusion power plant that generates more energy than is put into it; in other words, it produces a net output of power. Researchers worldwide are making progress toward meeting that goal.

The second challenge that Li cites sounds straightforward: “How do we get the heat out?” But understanding the problem and finding a solution are both far from obvious.

Research in the MIT Energy Initiative (MITEI) includes development and testing of advanced materials that may help address those challenges, as well as many other challenges of the energy transition. MITEI has multiple corporate members that have been supporting MIT’s efforts to advance technologies required to harness fusion energy.

The problem: An abundance of helium, a destructive force

Key to a fusion reactor is a superheated plasma — an ionized gas — that’s reacting inside a vacuum vessel. As light atoms in the plasma combine to form heavier ones, they release fast neutrons with high kinetic energy that shoot through the surrounding vacuum vessel into a coolant. During this process, those fast neutrons gradually lose their energy by causing radiation damage and generating heat. The heat that’s transferred to the coolant is eventually used to raise steam that drives an electricity-generating turbine.

The problem is finding a material for the vacuum vessel that remains strong enough to keep the reacting plasma and the coolant apart, while allowing the fast neutrons to pass through to the coolant. If one considers only the damage due to neutrons knocking atoms out of position in the metal structure, the vacuum vessel should last a full decade. However, depending on what materials are used in the fabrication of the vacuum vessel, some projections indicate that the vacuum vessel will last only six to 12 months. Why is that? Today’s nuclear fission reactors also generate neutrons, and those reactors last far longer than a year.

The difference is that fusion neutrons possess much higher kinetic energy than fission neutrons do, and as they penetrate the vacuum vessel walls, some of them interact with the nuclei of atoms in the structural material, giving off particles that rapidly turn into helium atoms. The result is hundreds of times more helium atoms than are present in a fission reactor. Those helium atoms look for somewhere to land — a place with low “embedding energy,” a measure that indicates how much energy it takes for a helium atom to be absorbed. As Li explains, “The helium atoms like to go to places with low helium embedding energy.” And in the metals used in fusion vacuum vessels, there are places with relatively low helium embedding energy — namely, naturally occurring openings called grain boundaries.

Metals are made up of individual grains inside which atoms are lined up in an orderly fashion. Where the grains come together there are gaps where the atoms don’t line up as well. That open space has relatively low helium embedding energy, so the helium atoms congregate there. Worse still, helium atoms have a repellent interaction with other atoms, so the helium atoms basically push open the grain boundary. Over time, the opening grows into a continuous crack, and the vacuum vessel breaks.

That congregation of helium atoms explains why the structure fails much sooner than expected based just on the number of helium atoms that are present. Li offers an analogy to illustrate. “Babylon is a city of a million people. But the claim is that 100 bad persons can destroy the whole city — if all those bad persons work at the city hall.” The solution? Give those bad persons other, more attractive places to go, ideally in their own villages.

To Li, the problem and possible solution are the same in a fusion reactor. If many helium atoms go to the grain boundary at once, they can destroy the metal wall. The solution? Add a small amount of a material that has a helium embedding energy even lower than that of the grain boundary. And over the past two years, Li and his team have demonstrated — both theoretically and experimentally — that their diversionary tactic works. By adding nanoscale particles of a carefully selected second material to the metal wall, they’ve found they can keep the helium atoms that form from congregating in the structurally vulnerable grain boundaries in the metal.

Looking for helium-absorbing compounds

To test their idea, So Yeon Kim ScD ’23 of the Department of Materials Science and Engineering and Haowei Xu PhD ’23 of the Department of Nuclear Science and Engineering acquired a sample composed of two materials, or “phases,” one with a lower helium embedding energy than the other. They and their collaborators then implanted helium ions into the sample at a temperature similar to that in a fusion reactor and watched as bubbles of helium formed. Transmission electron microscope images confirmed that the helium bubbles occurred predominantly in the phase with the lower helium embedding energy. As Li notes, “All the damage is in that phase — evidence that it protected the phase with the higher embedding energy.”

Having confirmed their approach, the researchers were ready to search for helium-absorbing compounds that would work well with iron, which is often the principal metal in vacuum vessel walls. “But calculating helium embedding energy for all sorts of different materials would be computationally demanding and expensive,” says Kim. “We wanted to find a metric that is easy to compute and a reliable indicator of helium embedding energy.”

They found such a metric: the “atomic-scale free volume,” which is basically the maximum size of the internal vacant space available for helium atoms to potentially settle. “This is just the radius of the largest sphere that can fit into a given crystal structure,” explains Kim. “It is a simple calculation.” Examination of a series of possible helium-absorbing ceramic materials confirmed that atomic free volume correlates well with helium embedding energy. Moreover, many of the ceramics they investigated have higher free volume, thus lower embedding energy, than the grain boundaries do.

However, in order to identify options for the nuclear fusion application, the screening needed to include some other factors. For example, in addition to the atomic free volume, a good second phase must be mechanically robust (able to sustain a load); it must not get very radioactive with neutron exposure; and it must be compatible — but not too cozy — with the surrounding metal, so it disperses well but does not dissolve into the metal. “We want to disperse the ceramic phase uniformly in the bulk metal to ensure that all grain boundary regions are close to the dispersed ceramic phase so it can provide protection to those regions,” says Li. “The two phases need to coexist, so the ceramic won’t either clump together or totally dissolve in the iron.”

Using their analytical tools, Kim and Xu examined about 50,000 compounds and identified 750 potential candidates. Of those, a good option for inclusion in a vacuum vessel wall made mainly of iron was iron silicate.

Experimental testing

The researchers were ready to examine samples in the lab. To make the composite material for proof-of-concept demonstrations, Kim and collaborators dispersed nanoscale particles of iron silicate into iron and implanted helium into that composite material. She took X-ray diffraction (XRD) images before and after implanting the helium and also computed the XRD patterns. The ratio between the implanted helium and the dispersed iron silicate was carefully controlled to allow a direct comparison between the experimental and computed XRD patterns. The measured XRD intensity changed with the helium implantation exactly as the calculations had predicted. “That agreement confirms that atomic helium is being stored within the bulk lattice of the iron silicate,” says Kim.

To follow up, Kim directly counted the number of helium bubbles in the composite. In iron samples without the iron silicate added, grain boundaries were flanked by many helium bubbles. In contrast, in the iron samples with the iron silicate ceramic phase added, helium bubbles were spread throughout the material, with many fewer occurring along the grain boundaries. Thus, the iron silicate had provided sites with low helium-embedding energy that lured the helium atoms away from the grain boundaries, protecting those vulnerable openings and preventing cracks from opening up and causing the vacuum vessel to fail catastrophically.

The researchers conclude that adding just 1 percent (by volume) of iron silicate to the iron walls of the vacuum vessel will cut the number of helium bubbles in half and also reduce their diameter by 20 percent — “and having a lot of small bubbles is OK if they’re not in the grain boundaries,” explains Li.

Next steps

Thus far, Li and his team have gone from computational studies of the problem and a possible solution to experimental demonstrations that confirm their approach. And they’re well on their way to commercial fabrication of components. “We’ve made powders that are compatible with existing commercial 3D printers and are preloaded with helium-absorbing ceramics,” says Li. The helium-absorbing nanoparticles are well dispersed and should provide sufficient helium uptake to protect the vulnerable grain boundaries in the structural metals of the vessel walls. While Li confirms that there’s more scientific and engineering work to be done, he, along with Alexander O'Brien PhD ’23 of the Department of Nuclear Science and Engineering and Kang Pyo So, a former postdoc in the same department, have already developed a startup company that’s ready to 3D print structural materials that can meet all the challenges faced by the vacuum vessel inside a fusion reactor.

This research was supported by Eni S.p.A. through the MIT Energy Initiative. Additional support was provided by a Kwajeong Scholarship; the U.S. Department of Energy (DOE) Laboratory Directed Research and Development program at Idaho National Laboratory; U.S. DOE Lawrence Livermore National Laboratory; and Creative Materials Discovery Program through the National Research Foundation of Korea.

A tiny battery designed by MIT engineers could enable the deployment of cell-sized, autonomous robots for drug delivery within in the human body, as well as other applications such as locating leaks in gas pipelines.

The new battery, which is 0.1 millimeters long and 0.002 millimeters thick — roughly the thickness of a human hair — can capture oxygen from air and use it to oxidize zinc, creating a current with a potential of up to 1 volt. That is enough to power a small circuit, sensor, or actuator, the researchers showed.

“We think this is going to be very enabling for robotics,” says Michael Strano, the Carbon P. Dubbs Professor of Chemical Engineering at MIT and the senior author of the study. “We’re building robotic functions onto the battery and starting to put these components together into devices.”

Ge Zhang PhD ’22 and Sungyun Yang, an MIT graduate student, are the lead author of the paper, which appears in Science Robotics.

Powered by batteries

For several years, Strano’s lab has been working on tiny robots that can sense and respond to stimuli in their environment. One of the major challenges in developing such tiny robots is making sure that they have enough power.

Other researchers have shown that they can power microscale devices using solar power, but the limitation to that approach is that the robots must have a laser or another light source pointed at them at all times. Such devices are known as “marionettes” because they are controlled by an external power source. Putting a power source such as a battery inside these tiny devices could free them to roam much farther.

“The marionette systems don’t really need a battery because they’re getting all the energy they need from outside,” Strano says. “But if you want a small robot to be able to get into spaces that you couldn’t access otherwise, it needs to have a greater level of autonomy. A battery is essential for something that’s not going to be tethered to the outside world.”

To create robots that could become more autonomous, Strano’s lab decided to use a type of battery known as a zinc-air battery. These batteries, which have a longer lifespan than many other types of batteries due to their high energy density, are often used in hearing aids.

The battery that they designed consists of a zinc electrode connected to a platinum electrode, embedded into a strip of a polymer called SU-8, which is commonly used for microelectronics. When these electrodes interact with oxygen molecules from the air, the zinc becomes oxidized and releases electrons that flow to the platinum electrode, creating a current.

In this study, the researchers showed that this battery could provide enough energy to power an actuator — in this case, a robotic arm that can be raised and lowered. The battery could also power a memristor, an electrical component that can store memories of events by changing its electrical resistance, and a clock circuit, which allows robotic devices to keep track of time.

The battery also provides enough power to run two different types of sensors that change their electrical resistance when they encounter chemicals in the environment. One of the sensors is made from atomically thin molybdenum disulfide and the other from carbon nanotubes.

“We’re making the basic building blocks in order to build up functions at the cellular level,” Strano says.

Robotic swarms

In this study, the researchers used a wire to connect their battery to an external device, but in future work they plan to build robots in which the battery is incorporated into a device.

“This is going to form the core of a lot of our robotic efforts,” Strano says. “You can build a robot around an energy source, sort of like you can build an electric car around the battery.”

One of those efforts revolves around designing tiny robots that could be injected into the human body, where they could seek out a target site and then release a drug such as insulin. For use in the human body, the researchers envision that the devices would be made of biocompatible materials that would break apart once they were no longer needed.

The researchers are also working on increasing the voltage of the battery, which may enable additional applications.

The research was funded by the U.S. Army Research Office, the U.S. Department of Energy, the National Science Foundation, and a MathWorks Engineering Fellowship.

In late 2023, the first drug with potential to slow the progression of Alzheimer's disease was approved by the U.S. Federal Drug Administration. Alzheimer's is one of many debilitating neurological disorders that together affect one-eighth of the world's population, and while the new drug is a step in the right direction, there is still a long journey ahead to fully understanding it, and other such diseases.

"Reconstructing the intricacies of how the human brain functions on a cellular level is one of the biggest challenges in neuroscience," says Lars Gjesteby, a technical staff member and algorithm developer from the MIT Lincoln Laboratory's Human Health and Performance Systems Group. "High-resolution, networked brain atlases can help improve our understanding of disorders by pinpointing differences between healthy and diseased brains. However, progress has been hindered by insufficient tools to visualize and process very large brain imaging datasets."

A networked brain atlas is in essence a detailed map of the brain that can help link structural information with neural function. To build such atlases, brain imaging data need to be processed and annotated. For example, each axon, or thin fiber connecting neurons, needs to be traced, measured, and labeled with information. Current methods of processing brain imaging data, such as desktop-based software or manual-oriented tools, are not yet designed to handle human brain-scale datasets. As such, researchers often spend a lot of time slogging through an ocean of raw data.

Gjesteby is leading a project to build the Neuron Tracing and Active Learning Environment (NeuroTrALE), a software pipeline that brings machine learning, supercomputing, as well as ease of use and access to this brain mapping challenge. NeuroTrALE automates much of the data processing and displays the output in an interactive interface that allows researchers to edit and manipulate the data to mark, filter, and search for specific patterns.

Untangling a ball of yarn

One of NeuroTrALE's defining features is the machine-learning technique it employs, called active learning. NeuroTrALE's algorithms are trained to automatically label incoming data based on existing brain imaging data, but unfamiliar data can present potential for errors. Active learning allows users to manually correct errors, teaching the algorithm to improve the next time it encounters similar data. This mix of automation and manual labeling ensures accurate data processing with a much smaller burden on the user.

"Imagine taking an X-ray of a ball of yarn. You'd see all these crisscrossed, overlapping lines," says Michael Snyder, from the laboratory's Homeland Decision Support Systems Group. "When two lines cross, does it mean one of the pieces of yarn is making a 90-degree bend, or is one going straight up and the other is going straight over? With NeuroTrALE's active learning, users can trace these strands of yarn one or two times and train the algorithm to follow them correctly moving forward. Without NeuroTrALE, the user would have to trace the ball of yarn, or in this case the axons of the human brain, every single time." Snyder is a software developer on the NeuroTrALE team along with staff member David Chavez.

Because NeuroTrALE takes the bulk of the labeling burden off of the user, it allows researchers to process more data more quickly. Further, the axon tracing algorithms harness parallel computing to distribute computations across multiple GPUs at once, leading to even faster, scalable processing. Using NeuroTrALE, the team demonstrated a 90 percent decrease in computing time needed to process 32 gigabytes of data over conventional AI methods.

The team also showed that a substantial increase in the volume of data does not translate to an equivalent increase in processing time. For example, in a recent study they demonstrated that a 10,000 percent increase in dataset size resulted in only a 9 percent and a 22 percent increase in total data processing time, using two different types of central processing units.

"With the estimated 86 billion neurons making 100 trillion connections in the human brain, manually labeling all the axons in a single brain would take lifetimes," adds Benjamin Roop, one of the project's algorithm developers. "This tool has the potential to automate the creation of connectomes for not just one individual, but many. That opens the door for studying brain disease at the population level."

The open-source road to discovery

The NeuroTrALE project was formed as an internally funded collaboration between Lincoln Laboratory and Professor Kwanghun Chung's laboratory on MIT campus. The Lincoln Lab team needed to build a way for the Chung Lab researchers to analyze and extract useful information from their large amount of brain imaging data flowing into the MIT SuperCloud — a supercomputer run by Lincoln Laboratory to support MIT research. Lincoln Lab's expertise in high-performance computing, image processing, and artificial intelligence made it exceptionally suited to tackling this challenge.

In 2020, the team uploaded NeuroTrALE to the SuperCloud and by 2022 the Chung Lab was producing results. In one study, published in Science, they used NeuroTrALE to quantify prefrontal cortex cell density in relation to Alzheimer's disease, where brains affected with the disease had a lower cell density in certain regions than those without. The same team also located where in the brain harmful neurofibers tend to get tangled in Alzheimer's-affected brain tissue.

Work on NeuroTrALE has continued with Lincoln Laboratory funding and funding from the National Institutes of Health (NIH) to build up NeuroTrALE's capabilities. Currently, its user interface tools are being integrated with Google's Neuroglancer program — an open-source, web-based viewer application for neuroscience data. NeuroTrALE adds the ability for users to visualize and edit their annotated data dynamically, and for multiple users to work with the same data at the same time. Users can also create and edit a number of shapes such as polygons, points, and lines to facilitate annotation tasks, as well as customize color display for each annotation to distinguish neurons in dense regions.

"NeuroTrALE provides a platform-agnostic, end-to-end solution that can be easily and rapidly deployed on standalone, virtual, cloud, and high performance computing environments via containers." says Adam Michaleas, a high performance computing engineer from the laboratory's Artificial Intelligence Technology Group. "Furthermore, it significantly improves the end user experience by providing capabilities for real-time collaboration within the neuroscience community via data visualization and simultaneous content review."

To align with NIH's mission of sharing research products, the team's goal is to make NeuroTrALE a fully open-source tool for anyone to use. And this type of tool, says Gjesteby, is what's needed to reach the end goal of mapping the entirety of the human brain for research, and eventually drug development. "It's a grassroots effort by the community where data and algorithms are meant to be shared and accessed by all."

The codebases for the axon tracing, data management, and interactive user interface of NeuroTrALE are publicly available via open-source licenses. Please contact Lars Gjesteby for more information on using NeuroTrALE. 

Ask a large language model (LLM) like GPT-4 to smell a rain-soaked campsite, and it’ll politely decline. Ask the same system to describe that scent to you, and it’ll wax poetic about “an air thick with anticipation" and “a scent that is both fresh and earthy," despite having neither prior experience with rain nor a nose to help it make such observations. One possible explanation for this phenomenon is that the LLM is simply mimicking the text present in its vast training data, rather than working with any real understanding of rain or smell.

But does the lack of eyes mean that language models can’t ever “understand" that a lion is “larger" than a house cat? Philosophers and scientists alike have long considered the ability to assign meaning to language a hallmark of human intelligence — and pondered what essential ingredients enable us to do so.

Peering into this enigma, researchers from MIT’s Computer Science and Artificial Intelligence Laboratory (CSAIL) have uncovered intriguing results suggesting that language models may develop their own understanding of reality as a way to improve their generative abilities. The team first developed a set of small Karel puzzles, which consisted of coming up with instructions to control a robot in a simulated environment. They then trained an LLM on the solutions, but without demonstrating how the solutions actually worked. Finally, using a machine learning technique called “probing,” they looked inside the model’s “thought process” as it generates new solutions. 

After training on over 1 million random puzzles, they found that the model spontaneously developed its own conception of the underlying simulation, despite never being exposed to this reality during training. Such findings call into question our intuitions about what types of information are necessary for learning linguistic meaning — and whether LLMs may someday understand language at a deeper level than they do today.

“At the start of these experiments, the language model generated random instructions that didn’t work. By the time we completed training, our language model generated correct instructions at a rate of 92.4 percent,” says MIT electrical engineering and computer science (EECS) PhD student and CSAIL affiliate Charles Jin, who is the lead author of a new paper on the work. “This was a very exciting moment for us because we thought that if your language model could complete a task with that level of accuracy, we might expect it to understand the meanings within the language as well. This gave us a starting point to explore whether LLMs do in fact understand text, and now we see that they’re capable of much more than just blindly stitching words together.”

Inside the mind of an LLM

The probe helped Jin witness this progress firsthand. Its role was to interpret what the LLM thought the instructions meant, unveiling that the LLM developed its own internal simulation of how the robot moves in response to each instruction. As the model’s ability to solve puzzles improved, these conceptions also became more accurate, indicating that the LLM was starting to understand the instructions. Before long, the model was consistently putting the pieces together correctly to form working instructions.

Jin notes that the LLM’s understanding of language develops in phases, much like how a child learns speech in multiple steps. Starting off, it’s like a baby babbling: repetitive and mostly unintelligible. Then, the language model acquires syntax, or the rules of the language. This enables it to generate instructions that might look like genuine solutions, but they still don’t work.

The LLM’s instructions gradually improve, though. Once the model acquires meaning, it starts to churn out instructions that correctly implement the requested specifications, like a child forming coherent sentences.

Separating the method from the model: A “Bizarro World”

The probe was only intended to “go inside the brain of an LLM” as Jin characterizes it, but there was a remote possibility that it also did some of the thinking for the model. The researchers wanted to ensure that their model understood the instructions independently of the probe, instead of the probe inferring the robot’s movements from the LLM’s grasp of syntax.

“Imagine you have a pile of data that encodes the LM’s thought process,” suggests Jin. “The probe is like a forensics analyst: You hand this pile of data to the analyst and say, ‘Here’s how the robot moves, now try and find the robot’s movements in the pile of data.’ The analyst later tells you that they know what’s going on with the robot in the pile of data. But what if the pile of data actually just encodes the raw instructions, and the analyst has figured out some clever way to extract the instructions and follow them accordingly? Then the language model hasn't really learned what the instructions mean at all.”

To disentangle their roles, the researchers flipped the meanings of the instructions for a new probe. In this “Bizarro World,” as Jin calls it, directions like “up” now meant “down” within the instructions moving the robot across its grid. 

“If the probe is translating instructions to robot positions, it should be able to translate the instructions according to the bizarro meanings equally well,” says Jin. “But if the probe is actually finding encodings of the original robot movements in the language model’s thought process, then it should struggle to extract the bizarro robot movements from the original thought process.”

As it turned out, the new probe experienced translation errors, unable to interpret a language model that had different meanings of the instructions. This meant the original semantics were embedded within the language model, indicating that the LLM understood what instructions were needed independently of the original probing classifier.

“This research directly targets a central question in modern artificial intelligence: are the surprising capabilities of large language models due simply to statistical correlations at scale, or do large language models develop a meaningful understanding of the reality that they are asked to work with? This research indicates that the LLM develops an internal model of the simulated reality, even though it was never trained to develop this model,” says Martin Rinard, an MIT professor in EECS, CSAIL member, and senior author on the paper.

This experiment further supported the team’s analysis that language models can develop a deeper understanding of language. Still, Jin acknowledges a few limitations to their paper: They used a very simple programming language and a relatively small model to glean their insights. In an upcoming work, they’ll look to use a more general setting. While Jin’s latest research doesn’t outline how to make the language model learn meaning faster, he believes future work can build on these insights to improve how language models are trained.

“An intriguing open question is whether the LLM is actually using its internal model of reality to reason about that reality as it solves the robot navigation problem,” says Rinard. “While our results are consistent with the LLM using the model in this way, our experiments are not designed to answer this next question.”

“There is a lot of debate these days about whether LLMs are actually ‘understanding’ language or rather if their success can be attributed to what is essentially tricks and heuristics that come from slurping up large volumes of text,” says Ellie Pavlick, assistant professor of computer science and linguistics at Brown University, who was not involved in the paper. “These questions lie at the heart of how we build AI and what we expect to be inherent possibilities or limitations of our technology. This is a nice paper that looks at this question in a controlled way — the authors exploit the fact that computer code, like natural language, has both syntax and semantics, but unlike natural language, the semantics can be directly observed and manipulated for experimental purposes. The experimental design is elegant, and their findings are optimistic, suggesting that maybe LLMs can learn something deeper about what language ‘means.’”

Jin and Rinard’s paper was supported, in part, by grants from the U.S. Defense Advanced Research Projects Agency (DARPA). 

In 2023, more than 100,000 Americans died from opioid overdoses. The most effective way to save someone who has overdosed is to administer a drug called naloxone, but a first responder or bystander can’t always reach the person who has overdosed in time.

Researchers at MIT and Brigham and Women’s Hospital have developed a new device that they hope will help to eliminate those delays and potentially save the lives of people who overdose. The device, about the size of a stick of gum, can be implanted under the skin, where it monitors heart rate, breathing rate, and other vital signs. When it determines that an overdose has occurred, it rapidly pumps out a dose of naloxone.

In a study appearing today in the journal Device, the researchers showed that the device can successfully reverse overdoses in animals. With further development, the researchers envision that this approach could provide a new option for helping to prevent overdose deaths in high-risk populations, such as people who have already survived an overdose.

“This could really address a significant unmet need in the population that suffers from substance abuse and opiate dependency to help mitigate overdoses, with the initial focus on the high-risk population,” says Giovanni Traverso, an associate professor of mechanical engineering at MIT, a gastroenterologist at Brigham and Women’s Hospital, and the senior author of the study.

The paper’s lead authors are Hen-Wei Huang, a former MIT visiting scientist and currently an assistant professor of electrical and electronic engineering at Nanyang Technological University in Singapore; Peter Chai, an associate professor of emergency medicine physician at Brigham and Women’s Hospital; SeungHo Lee, a research scientist at MIT’s Koch Institute for Integrative Cancer Research; Tom Kerssemakers and Ali Imani, former master’s students at Brigham and Women’s Hospital; and Jack Chen, a doctoral student in mechanical engineering at MIT.

An implantable device

Naloxone is an opioid antagonist, meaning that it can bind to opioid receptors and block the effects of other opioids, including heroin and fentanyl. The drug, which is given by injection or as a nasal spray, can restore normal breathing within just a few minutes of being administered.

However, many people are alone when they overdose, and may not receive assistance in time to save their lives. Additionally, with a new wave of synthetic, more potent opioids sweeping the U.S., opioid overdoses can be more rapid in onset and unpredictable. To try to overcome that, some researchers are developing wearable devices that could detect an overdose and administer naloxone, but none of those have yet proven successful. The MIT/BWH team set out to design an implantable device that would be less bulky, provide direct injection of naloxone into the subcutaneous tissue, and eliminate the need for the patient to remember to wear it.

The device that the researchers came up with includes sensors that can detect heart rate, breathing rate, blood pressure, and oxygen saturation. In an animal study, the researchers used the sensors to measure all of these signals and determine exactly how they change during an overdose of fentanyl. This resulted in a unique algorithm that increases the sensitivity of the device to accurately detect opioid overdose and distinguish it from other conditions where breathing is decreased, such as sleep apnea.

This study showed that fentanyl first leads to a drop in heart rate, followed quickly by a slowdown of breathing. By measuring how these signals changed, the researchers were able to calculate the point at which naloxone administration should be triggered.

“The most challenging aspect of developing an engineering solution to prevent overdose mortality is simultaneously addressing patient adherence and willingness to adopt new technology, combating stigma, minimizing false positive detections, and ensuring the rapid delivery of antidotes,” says Huang. “Our proposed solution tackles these unmet needs by developing a miniaturized robotic implant equipped with multisensing modalities, continuous monitoring capabilities, on-board decision making, and an innovative micropumping mechanism.”

The device also includes a small reservoir that can carry up to 10 milligrams of naloxone. When an overdose is detected, it triggers a pump that ejects the naloxone, which is released within about 10 seconds.

In their animal studies, the researchers found that this drug administration could reverse the effects of an overdose 96 percent of the time.

“We created a closed-loop system that can sense the onset of the opiate overdose and then release the antidote, and then you see that recovery,” Traverso says.

Preventing overdoses

The researchers envision that this technology could be used to help people who are at the highest risk of overdose, beginning with people who have had a previous overdose. They now plan to investigate how to make the device as user-friendly as possible, studying factors such as the optimal location for implantation.

“A key pillar of addressing the opioid epidemic is providing naloxone to individuals at key moments of risk. Our vision for this device is for it to integrate into the cascade of harm-reduction strategies to efficiently and safely deliver naloxone, preventing death from opioid overdose and providing the opportunity to support individuals with opioid use disorder,” says Chai.

The researchers hope to be able to test the device in humans within the next three to five years. They are now working on miniaturizing the device further and optimizing the on-board battery, which currently can provide power for about two weeks.

The research was funded by Novo Nordisk, the McGraw Family Foundation at Brigham and Women’s Hospital, and the MIT Department of Mechanical Engineering.

In a new study appearing today in the journal AGU Advances, scientists at MIT and NASA report that seven rock samples collected along the “fan front” of Mars’ Jezero Crater contain minerals that are typically formed in water. The findings suggest that the rocks were originally deposited by water, or may have formed in the presence of water.

The seven samples were collected by NASA’s Perseverance rover in 2022 during its exploration of the crater’s western slope, where some rocks were hypothesized to have formed in what is now a dried-up ancient lake. Members of the Perseverance science team, including MIT scientists, have studied the rover’s images and chemical analyses of the samples, and confirmed that the rocks indeed contain signs of water, and that the crater was likely once a watery, habitable environment.

Whether the crater was actually inhabited is yet unknown. The team found that the presence of organic matter — the starting material for life — cannot be confirmed, at least based on the rover’s measurements. But judging from the rocks’ mineral content, scientists believe the samples are their best chance of finding signs of ancient Martian life once the rocks are returned to Earth for more detailed analysis.

“These rocks confirm the presence, at least temporarily, of habitable environments on Mars,” says the study’s lead author, Tanja Bosak, professor of geobiology in MIT’s Department of Earth, Atmospheric and Planetary Sciences (EAPS). “What we’ve found is that indeed there was a lot of water activity. For how long, we don’t know, but certainly for long enough to create these big sedimentary deposits.”

What’s more, some of the collected samples may have originally been deposited in the ancient lake more than 3.5 billion years ago — before even the first signs of life on Earth.

“These are the oldest rocks that may have been deposited by water, that we’ve ever laid hands or rover arms on,” says co-author Benjamin Weiss, the Robert R. Shrock Professor of Earth and Planetary Sciences at MIT. “That’s exciting, because it means these are the most promising rocks that may have preserved fossils, and signatures of life.”

The study’s MIT co-authors include postdoc Eva Scheller, and research scientist Elias Mansbach, along with members of the Perseverance science team.

At the front

The new rock samples were collected in 2022 as part of the rover’s Fan Front Campaign — an exploratory phase during which Perseverance traversed Jezero Crater’s western slope, where a fan-like region contains sedimentary, layered rocks. Scientists suspect that this “fan front” is an ancient delta that was created by sediment that flowed with a river and settled into a now bone-dry lakebed. If life existed on Mars, scientists believe that it could be preserved in the layers of sediment along the fan front.

In the end, Perseverance collected seven samples from various locations along the fan front. The rover obtained each sample by drilling into the Martian bedrock and extracting a pencil-sized core, which it then sealed in a tube to one day be retrieved and returned to Earth for detailed analysis.

Prior to extracting the cores, the rover took images of the surrounding sediments at each of the seven locations. The science team then processed the imaging data to estimate a sediment’s average grain size and mineral composition. This analysis showed that all seven collected samples likely contain signs of water, suggesting that they were initially deposited by water.

Specifically, Bosak and her colleagues found evidence of certain minerals in the sediments that are known to precipitate out of water.

“We found lots of minerals like carbonates, which are what make reefs on Earth,” Bosak says. “And it’s really an ideal material that can preserve fossils of microbial life.”

Interestingly, the researchers also identified sulfates in some samples that were collected at the base of the fan front. Sulfates are minerals that form in very salty water — another sign that water was present in the crater at one time — though very salty water, Bosak notes, “is not necessarily the best thing for life.” If the entire crater was once filled with very salty water, then it would be difficult for any form of life to thrive. But if only the bottom of the lake were briny, that could be an advantage, at least for preserving any signs of life that may have lived further up, in less salty layers, that eventually died and drifted down to the bottom.

“However salty it was, if there were any organics present, it's like pickling something in salt,” Bosak says. “If there was life that fell into the salty layer, it would be very well-preserved.”

Fuzzy fingerprints

But the team emphasizes that organic matter has not been confidently detected by the rover’s instruments. Organic matter can be signs of life, but can also be produced by certain geological processes that have nothing to do with living matter. Perseverance’s predecessor, the Curiosity rover, had detected organic matter throughout Mars’ Gale Crater, which scientists suspect may have come from asteroids that made impact with Mars in the past.

And in a previous campaign, Perseverance detected what appeared to be organic molecules at multiple locations along Jezero Crater’s floor. These observations were taken by the rover’s Scanning Habitable Environments with Raman and Luminescence for Organics and Chemicals (SHERLOC) instrument, which uses ultraviolet light to scan the Martian surface. If organics are present, they can glow, similar to material under a blacklight. The wavelengths at which the material glows act as a sort of fingerprint for the kind of organic molecules that are present.

In Perseverance’s previous exploration of the crater floor, SHERLOC appeared to pick up signs of organic molecules throughout the region, and later, at some locations along the fan front. But a careful analysis, led by MIT’s Eva Scheller, has found that while the particular wavelengths observed could be signs of organic matter, they could just as well be signatures of substances that have nothing to do with organic matter.

“It turns out that cerium metals incorporated in minerals actually produce very similar signals as the organic matter,” Scheller says. “When investigated, the potential organic signals were strongly correlated with phosphate minerals, which always contain some cerium.”

Scheller’s work shows that the rover’s measurements cannot be interpreted definitively as organic matter.

“This is not bad news,” Bosak says. “It just tells us there is not very abundant organic matter. It’s still possible that it’s there. It’s just below the rover’s detection limit.”

When the collected samples are finally sent back to Earth, Bosak says laboratory instruments will have more than enough sensitivity to detect any organic matter that might lie within.

“On Earth, once we have microscopes with nanometer-scale resolution, and various types of instruments that we cannot staff on one rover, then we can actually attempt to look for life,” she says.

This work was supported, in part, by NASA.

Identifying one faulty turbine in a wind farm, which can involve looking at hundreds of signals and millions of data points, is akin to finding a needle in a haystack.

Engineers often streamline this complex problem using deep-learning models that can detect anomalies in measurements taken repeatedly over time by each turbine, known as time-series data.

But with hundreds of wind turbines recording dozens of signals each hour, training a deep-learning model to analyze time-series data is costly and cumbersome. This is compounded by the fact that the model may need to be retrained after deployment, and wind farm operators may lack the necessary machine-learning expertise.

In a new study, MIT researchers found that large language models (LLMs) hold the potential to be more efficient anomaly detectors for time-series data. Importantly, these pretrained models can be deployed right out of the box.

The researchers developed a framework, called SigLLM, which includes a component that converts time-series data into text-based inputs an LLM can process. A user can feed these prepared data to the model and ask it to start identifying anomalies. The LLM can also be used to forecast future time-series data points as part of an anomaly detection pipeline.

While LLMs could not beat state-of-the-art deep learning models at anomaly detection, they did perform as well as some other AI approaches. If researchers can improve the performance of LLMs, this framework could help technicians flag potential problems in equipment like heavy machinery or satellites before they occur, without the need to train an expensive deep-learning model.

“Since this is just the first iteration, we didn’t expect to get there from the first go, but these results show that there’s an opportunity here to leverage LLMs for complex anomaly detection tasks,” says Sarah Alnegheimish, an electrical engineering and computer science (EECS) graduate student and lead author of a paper on SigLLM.

Her co-authors include Linh Nguyen, an EECS graduate student; Laure Berti-Equille, a research director at the French National Research Institute for Sustainable Development; and senior author Kalyan Veeramachaneni, a principal research scientist in the Laboratory for Information and Decision Systems. The research will be presented at the IEEE Conference on Data Science and Advanced Analytics.

An off-the-shelf solution

Large language models are autoregressive, which means they can understand that the newest values in sequential data depend on previous values. For instance, models like GPT-4 can predict the next word in a sentence using the words that precede it.

Since time-series data are sequential, the researchers thought the autoregressive nature of LLMs might make them well-suited for detecting anomalies in this type of data.

However, they wanted to develop a technique that avoids fine-tuning, a process in which engineers retrain a general-purpose LLM on a small amount of task-specific data to make it an expert at one task. Instead, the researchers deploy an LLM off the shelf, with no additional training steps.

But before they could deploy it, they had to convert time-series data into text-based inputs the language model could handle.

They accomplished this through a sequence of transformations that capture the most important parts of the time series while representing data with the fewest number of tokens. Tokens are the basic inputs for an LLM, and more tokens require more computation.

“If you don’t handle these steps very carefully, you might end up chopping off some part of your data that does matter, losing that information,” Alnegheimish says.

Once they had figured out how to transform time-series data, the researchers developed two anomaly detection approaches.

Approaches for anomaly detection

For the first, which they call Prompter, they feed the prepared data into the model and prompt it to locate anomalous values.

“We had to iterate a number of times to figure out the right prompts for one specific time series. It is not easy to understand how these LLMs ingest and process the data,” Alnegheimish adds.

For the second approach, called Detector, they use the LLM as a forecaster to predict the next value from a time series. The researchers compare the predicted value to the actual value. A large discrepancy suggests that the real value is likely an anomaly.

With Detector, the LLM would be part of an anomaly detection pipeline, while Prompter would complete the task on its own. In practice, Detector performed better than Prompter, which generated many false positives.

“I think, with the Prompter approach, we were asking the LLM to jump through too many hoops. We were giving it a harder problem to solve,” says Veeramachaneni.

When they compared both approaches to current techniques, Detector outperformed transformer-based AI models on seven of the 11 datasets they evaluated, even though the LLM required no training or fine-tuning.

In the future, an LLM may also be able to provide plain language explanations with its predictions, so an operator could be better able to understand why an LLM identified a certain data point as anomalous.

However, state-of-the-art deep learning models outperformed LLMs by a wide margin, showing that there is still work to do before an LLM could be used for anomaly detection.

“What will it take to get to the point where it is doing as well as these state-of-the-art models? That is the million-dollar question staring at us right now. An LLM-based anomaly detector needs to be a game-changer for us to justify this sort of effort,” Veeramachaneni says.

Moving forward, the researchers want to see if finetuning can improve performance, though that would require additional time, cost, and expertise for training.

Their LLM approaches also take between 30 minutes and two hours to produce results, so increasing the speed is a key area of future work. The researchers also want to probe LLMs to understand how they perform anomaly detection, in the hopes of finding a way to boost their performance.

“When it comes to complex tasks like anomaly detection in time series, LLMs really are a contender. Maybe other complex tasks can be addressed with LLMs, as well?” says Alnegheimish.

This research was supported by SES S.A., Iberdrola and ScottishPower Renewables, and Hyundai Motor Company.

Early-stage trials in Alzheimer’s disease patients and studies in mouse models of the disease have suggested positive impacts on pathology and symptoms from exposure to light and sound presented at the “gamma” band frequency of 40 hertz (Hz). A new study zeroes in on how 40Hz sensory stimulation helps to sustain an essential process in which the signal-sending branches of neurons, called axons, are wrapped in a fatty insulation called myelin. Often called the brain’s “white matter,” myelin protects axons and insures better electrical signal transmission in brain circuits.

“Previous publications from our lab have mainly focused on neuronal protection,” says Li-Huei Tsai, Picower Professor in The Picower Institute for Learning and Memory and the Department of Brain and Cognitive Sciences at MIT and senior author of the new open-access study in Nature Communications. Tsai also leads MIT’s Aging Brain Initiative. “But this study shows that it’s not just the gray matter, but also the white matter that’s protected by this method.”

This year Cognito Therapeutics, the spinoff company that licensed MIT’s sensory stimulation technology, published phase II human trial results in the Journal of Alzheimer’s Disease indicating that 40Hz light and sound stimulation significantly slowed the loss of myelin in volunteers with Alzheimer’s. Also this year, Tsai’s lab published a study showing that gamma sensory stimulation helped mice withstand neurological effects of chemotherapy medicines, including by preserving myelin. In the new study, members of Tsai’s lab led by former postdoc Daniela Rodrigues Amorim used a common mouse model of myelin loss — a diet with the chemical cuprizone — to explore how sensory stimulation preserves myelination.

Amorim and Tsai’s team found that 40Hz light and sound not only preserved myelination in the brains of cuprizone-exposed mice, it also appeared to protect oligodendrocytes (the cells that myelinate neural axons), sustain the electrical performance of neurons, and preserve a key marker of axon structural integrity. When the team looked into the molecular underpinnings of these benefits, they found clear signs of specific mechanisms including preservation of neural circuit connections called synapses; a reduction in a cause of oligodendrocyte death called “ferroptosis;” reduced inflammation; and an increase in the ability of microglia brain cells to clean up myelin damage so that new myelin could be restored.

“Gamma stimulation promotes a healthy environment,” says Amorim, who is now a Marie Curie Fellow at the University of Galway in Ireland. “There are several ways we are seeing different effects.”

The findings suggest that gamma sensory stimulation may help not only Alzheimer’s disease patients but also people battling other diseases involving myelin loss, such as multiple sclerosis, the authors wrote in the study.

Maintaining myelin

To conduct the study, Tsai and Amorim’s team fed some male mice a diet with cuprizone and gave other male mice a normal diet for six weeks. Halfway into that period, when cuprizone is known to begin causing its most acute effects on myelination, they exposed some mice from each group to gamma sensory stimulation for the remaining three weeks. In this way they had four groups: completely unaffected mice, mice that received no cuprizone but did get gamma stimulation, mice that received cuprizone and constant (but not 40Hz) light and sound as a control, and mice that received cuprizone and also gamma stimulation.

After the six weeks elapsed, the scientists measured signs of myelination throughout the brains of the mice in each group. Mice that weren’t fed cuprizone maintained healthy levels, as expected. Mice that were fed cuprizone and didn’t receive 40Hz gamma sensory stimulation showed drastic levels of myelin loss. Cuprizone-fed mice that received 40Hz stimulation retained significantly more myelin, rivaling the health of mice never fed cuprizone by some, but not all, measures.

The researchers also looked at numbers of oligodendrocytes to see if they survived better with sensory stimulation. Several measures revealed that in mice fed cuprizone, oligodendrocytes in the corpus callosum region of the brain (a key point for the transit of neural signals because it connects the brain’s hemispheres) were markedly reduced. But in mice fed cuprizone and also treated with gamma stimulation, the number of cells were much closer to healthy levels.

Electrophysiological tests among neural axons in the corpus callosum showed that gamma sensory stimulation was associated with improved electrical performance in cuprizone-fed mice who received gamma stimulation compared to cuprizone-fed mice left untreated by 40Hz stimulation. And when researchers looked in the anterior cingulate cortex region of the brain, they saw that MAP2, a protein that signals the structural integrity of axons, was much better preserved in mice that received cuprizone and gamma stimulation compared to cuprizone-fed mice who did not.

A key goal of the study was to identify possible ways in which 40Hz sensory stimulation may protect myelin.

To find out, the researchers conducted a sweeping assessment of protein expression in each mouse group and identified which proteins were differentially expressed based on cuprizone diet and exposure to gamma frequency stimulation. The analysis revealed distinct sets of effects between the cuprizone mice exposed to control stimulation and cuprizone-plus-gamma mice.

A highlight of one set of effects was the increase in MAP2 in gamma-treated cuprizone-fed mice. A highlight of another set was that cuprizone mice who received control stimulation showed a substantial deficit in expression of proteins associated with synapses. The gamma-treated cuprizone-fed mice did not show any significant loss, mirroring results in a 2019 Alzheimer’s 40Hz study that showed synaptic preservation. This result is important, the researchers wrote, because neural circuit activity, which depends on maintaining synapses, is associated with preserving myelin. They confirmed the protein expression results by looking directly at brain tissues.

Another set of protein expression results hinted at another important mechanism: ferroptosis. This phenomenon, in which errant metabolism of iron leads to a lethal buildup of reactive oxygen species in cells, is a known problem for oligodendrocytes in the cuprizone mouse model. Among the signs was an increase in cuprizone-fed, control stimulation mice in expression of the protein HMGB1, which is a marker of ferroptosis-associated damage that triggers an inflammatory response. Gamma stimulation, however, reduced levels of HMGB1.

Looking more deeply at the cellular and molecular response to cuprizone demyelination and the effects of gamma stimulation, the team assessed gene expression using single-cell RNA sequencing technology. They found that astrocytes and microglia became very inflammatory in cuprizone-control mice but gamma stimulation calmed that response. Fewer cells became inflammatory and direct observations of tissue showed that microglia became more proficient at clearing away myelin debris, a key step in effecting repairs.

The team also learned more about how oligodendrocytes in cuprizone-fed mice exposed to 40Hz sensory stimulation managed to survive better. Expression of protective proteins such as HSP70 increased and as did expression of GPX4, a master regulator of processes that constrain ferroptosis.

In addition to Amorim and Tsai, the paper’s other authors are Lorenzo Bozzelli, TaeHyun Kim, Liwang Liu, Oliver Gibson, Cheng-Yi Yang, Mitch Murdock, Fabiola Galiana-Meléndez, Brooke Schatz, Alexis Davison, Md Rezaul Islam, Dong Shin Park, Ravikiran M. Raju, Fatema Abdurrob, Alissa J. Nelson, Jian Min Ren, Vicky Yang and Matthew P. Stokes.

Fundacion Bancaria la Caixa, The JPB Foundation, The Picower Institute for Learning and Memory, the Carol and Gene Ludwig Family Foundation, Lester A. Gimpelson, Eduardo Eurnekian, The Dolby Family, Kathy and Miguel Octavio, the Marc Haas Foundation, Ben Lenail and Laurie Yoler, and the U.S. National Institutes of Health provided funding for the study.

Quantum materials — those with electronic properties that are governed by the principles of quantum mechanics, such as correlation and entanglement — can exhibit exotic behaviors under certain conditions, such as the ability to transmit electricity without resistance, known as superconductivity. However, in order to get the best performance out of these materials, they need to be properly tuned, in the same way that race cars require tuning as well. A team led by Mingda Li, an associate professor in MIT’s Department of Nuclear Science and Engineering (NSE), has demonstrated a new, ultra-precise way to tweak the characteristics of quantum materials, using a particular class of these materials, Weyl semimetals, as an example.

The new technique is not limited to Weyl semimetals. “We can use this method for any inorganic bulk material, and for thin films as well,” maintains NSE postdoc Manasi Mandal, one of two lead authors of an open-access paper — published recently in Applied Physics Reviews — that reported on the group’s findings.

The experiment described in the paper focused on a specific type of Weyl semimetal, a tantalum phosphide (TaP) crystal. Materials can be classified by their electrical properties: metals conduct electricity readily, whereas insulators impede the free flow of electrons. A semimetal lies somewhere in between. It can conduct electricity, but only in a narrow frequency band or channel. Weyl semimetals are part of a wider category of so-called topological materials that have certain distinctive features. For instance, they possess curious electronic structures — kinks or “singularities” called Weyl nodes, which are swirling patterns around a single point (configured in either a clockwise or counterclockwise direction) that resemble hair whorls or, more generally, vortices. The presence of Weyl nodes confers unusual, as well as useful, electrical properties. And a key advantage of topological materials is that their sought-after qualities can be preserved, or “topologically protected,” even when the material is disturbed.

“That’s a nice feature to have,” explains Abhijatmedhi Chotrattanapituk, a PhD student in MIT’s Department of Electrical Engineering and Computer Science and the other lead author of the paper. “When you try to fabricate this kind of material, you don’t have to be exact. You can tolerate some imperfections, some level of uncertainty, and the material will still behave as expected.”

Like water in a dam

The “tuning” that needs to happen relates primarily to the Fermi level, which is the highest energy level occupied by electrons in a given physical system or material. Mandal and Chotrattanapituk suggest the following analogy: Consider a dam that can be filled with varying levels of water. One can raise that level by adding water or lower it by removing water. In the same way, one can adjust the Fermi level of a given material simply by adding or subtracting electrons.

To fine-tune the Fermi level of the Weyl semimetal, Li’s team did something similar, but instead of adding actual electrons, they added negative hydrogen ions (each consisting of a proton and two electrons) to the sample. The process of introducing a foreign particle, or defect, into the TaP crystal — in this case by substituting a hydrogen ion for a tantalum atom — is called doping. And when optimal doping is achieved, the Fermi level will coincide with the energy level of the Weyl nodes. That’s when the material’s desired quantum properties will be most fully realized.

For Weyl semimetals, the Fermi level is especially sensitive to doping. Unless that level is set close to the Weyl nodes, the material’s properties can diverge significantly from the ideal. The reason for this extreme sensitivity owes to the peculiar geometry of the Weyl node. If one were to think of the Fermi level as the water level in a reservoir, the reservoir in a Weyl semimetal is not shaped like a cylinder; it’s shaped like an hourglass, and the Weyl node is located at the narrowest point, or neck, of that hourglass. Adding too much or too little water would miss the neck entirely, just as adding too many or too few electrons to the semimetal would miss the node altogether.

Fire up the hydrogen

To reach the necessary precision, the researchers utilized MIT’s two-stage “Tandem” ion accelerator — located at the Center for Science and Technology with Accelerators and Radiation (CSTAR) — and buffeted the TaP sample with high-energy ions coming out of the powerful (1.7 million volt) accelerator beam. Hydrogen ions were chosen for this purpose because they are the smallest negative ions available and thus alter the material less than a much larger dopant would. “The use of advanced accelerator techniques allows for greater precision than was ever before possible, setting the Fermi level to milli-electron volt [thousandths of an electron volt] accuracy,” says Kevin Woller, the principal research scientist who leads the CSTAR lab. “Additionally, high-energy beams allow for the doping of bulk crystals beyond the limitations of thin films only a few tens of nanometers thick.”

The procedure, in other words, involves bombarding the sample with hydrogen ions until a sufficient number of electrons are taken in to make the Fermi level just right. The question is: how long do you run the accelerator, and how do you know when enough is enough? The point being that you want to tune the material until the Fermi level is neither too low nor too high.

“The longer you run the machine, the higher the Fermi level gets,” Chotrattanapituk says. “The difficulty is that we cannot measure the Fermi level while the sample is in the accelerator chamber.” The normal way to handle that would be to irradiate the sample for a certain amount of time, take it out, measure it, and then put it back in if the Fermi level is not high enough. “That can be practically impossible,” Mandal adds.

To streamline the protocol, the team has devised a theoretical model that first predicts how many electrons are needed to increase the Fermi level to the preferred level and translates that to the number of negative hydrogen ions that must be added to the sample. The model can then tell them how long the sample ought to be kept in the accelerator chamber.

The good news, Chotrattanapituk says, is that their simple model agrees within a factor of 2 with trusted conventional models that are much more computationally intensive and may require access to a supercomputer. The group’s main contributions are two-fold, he notes: offering a new, accelerator-based technique for precision doping and providing a theoretical model that can guide the experiment, telling researchers how much hydrogen should be added to the sample depending on the energy of the ion beam, the exposure time, and the size and thickness of the sample.

Fine things to come with fine-tuning

This could pave the way to a major practical advance, Mandal notes, because their approach can potentially bring the Fermi level of a sample to the requisite value in a matter of minutes — a task that, by conventional methods, has sometimes taken weeks without ever reaching the required degree of milli-eV precision.

Li believes that an accurate and convenient method for fine-tuning the Fermi level could have broad applicability. “When it comes to quantum materials, the Fermi level is practically everything,” he says. “Many of the effects and behaviors that we seek only manifest themselves when the Fermi level is at the right location.” With a well-adjusted Fermi level, for example, one could raise the critical temperature at which materials become superconducting. Thermoelectric materials, which convert temperature differences into an electrical voltage, similarly become more efficient when the Fermi level is set just right. Precision tuning might also play a helpful role in quantum computing.

Thomas Zac Ward, a senior scientist at the Oak Ridge National Laboratory, offered a bullish assessment: “This work provides a new route for the experimental exploration of the critical, yet still poorly understand, behaviors of emerging materials. The ability to precisely control the Fermi level of a topological material is an important milestone that can help bring new quantum information and microelectronics device architectures to fruition.”

MIT chemists have developed a new way to synthesize complex molecules that were originally isolated from plants and could hold potential as antibiotics, analgesics, or cancer drugs.

These compounds, known as oligocyclotryptamines, consist of multiple tricyclic substructures called cyclotryptamine, fused together by carbon–carbon bonds. Only small quantities of these compounds are naturally available, and synthesizing them in the lab has proven difficult. The MIT team came up with a way to add tryptamine-derived components to a molecule one at a time, in a way that allows the researchers to precisely assemble the rings and control the 3D orientation of each component as well as the final product.

“For many of these compounds, there hasn’t been enough material to do a thorough review of their potential. I’m hopeful that having access to these compounds in a reliable way will enable us to do further studies,” says Mohammad Movassaghi, an MIT professor of chemistry and the senior author of the new study.

In addition to allowing scientists to synthesize oligocyclotryptamines found in plants, this approach could also be used to generate new variants that may have even better medicinal properties, or molecular probes that can help to reveal their mechanism of action.

Tony Scott PhD ’23 is the lead author of the paper, which appears today in the Journal of the American Chemical Society.

Fusing rings

Oligocyclotryptamines belong to a class of molecules called alkaloids — nitrogen-containing organic compounds produced mainly by plants. At least eight different oligocyclotryptamines have been isolated from a genus of flowering plants known as Psychotria, most of which are found in tropical forests.

Since the 1950s, scientists have studied the structure and synthesis of dimeric cyclotryptamines, which have two cyclotryptamine subunits. Over the past 20 years, significant progress has been made characterizing and synthesizing dimers and other smaller members of the family. However, no one has been able to synthesize the largest oligocyclotryptamines, which have six or seven rings fused together.

One of the hurdles in synthesizing these molecules is a step that requires formation of a bond between a carbon atom of one tryptamine-derived subunit to a carbon atom of the next subunit. The oligocyclotryptamines have two types of these linkages, both containing at least one carbon atom that has bonds with four other carbons. That extra bulk makes those carbon atoms less accessible to undergo reactions, and controlling the stereochemistry — the orientation of the atoms around the carbon — at all these junctures poses a significant challenge.

For many years, Movassaghi’s lab has been developing ways to form carbon-carbon bonds between carbon atoms that are already crowded with other atoms. In 2011, they devised a method that involves transforming the two carbon atoms into carbon radicals (carbon atoms with one unpaired electron) and directing their union. To create these radicals, and guide the paired union to be completely selective, the researchers first attach each of the targeted carbon atoms to a nitrogen atom; these two nitrogen atoms bind to each other.

When the researchers shine certain wavelengths of light on the substrate containing the two fragments linked via the two nitrogen atoms, it causes the two atoms of nitrogen to break away as nitrogen gas, leaving behind two very reactive carbon radicals in close proximity that join together almost immediately. This type of bond formation has also allowed the researchers to control the molecules’ stereochemistry.

Movassaghi demonstrated this approach, which he calls diazene-directed assembly, by synthesizing other types of alkaloids, including the communesins. These compounds are found in fungi and consist of two ring-containing molecules, or monomers, joined together. Later, Movassaghi began using this approach to fuse larger numbers of monomers, and he and Scott eventually turned their attention to the largest oligocyclotryptamine alkaloids.

The synthesis that they developed begins with one molecule of cyclotryptamine derivative, to which additional cyclotryptamine fragments with correct relative stereochemistry and position selectivity are added, one at a time. Each of these additions is made possible by the diazene-directed process that Movassaghi’s lab previously developed.

“The reason why we’re excited about this is that this single solution allowed us to go after multiple targets,” Movassaghi says. “That same route provides us a solution to multiple members of the natural product family because by extending the iteration one more cycle, your solution is now applied to a new natural product.”

“A tour de force”

Using this approach, the researchers were able to create molecules with six or seven cyclotryptamine rings, which has never been done before.

“Researchers worldwide have been trying to find a way to make these molecules, and Movassaghi and Scott are the first to pull it off,” says Seth Herzon, a professor of chemistry at Yale University, who was not involved in the research. Herzon described the work as “a tour de force in organic synthesis.”

Now that the researchers have synthesized these naturally occurring oligocyclotryptamines, they should be able to generate enough of the compounds that their potential therapeutic activity can be more thoroughly investigated.

They should also be able to create novel compounds by switching in slightly different cyclotryptamine subunits, Movassaghi says.

“We will continue to use this very precise way of adding these cyclotryptamine units to assemble them together into complex systems that have not been addressed yet, including derivatives that could potentially have improved properties,” he says.

The research was funded by the U.S. National Institute of General Medical Sciences.

Pick-and-place machines are a type of automated equipment used to place objects into structured, organized locations. These machines are used for a variety of applications — from electronics assembly to packaging, bin picking, and even inspection — but many current pick-and-place solutions are limited. Current solutions lack “precise generalization,” or the ability to solve many tasks without compromising on accuracy.

“In industry, you often see that [manufacturers] end up with very tailored solutions to the particular problem that they have, so a lot of engineering and not so much flexibility in terms of the solution,” Maria Bauza Villalonga PhD ’22, a senior research scientist at Google DeepMind where she works on robotics and robotic manipulation. “SimPLE solves this problem and provides a solution to pick-and-place that is flexible and still provides the needed precision.”

A new paper by MechE researchers published in the journal Science Robotics explores pick-and-place solutions with more precision. In precise pick-and-place, also known as kitting, the robot transforms an unstructured arrangement of objects into an organized arrangement. The approach, dubbed SimPLE (Simulation to Pick Localize and placE), learns to pick, regrasp and place objects using the object’s computer-aided design (CAD) model, and all without any prior experience or encounters with the specific objects.

“The promise of SimPLE is that we can solve many different tasks with the same hardware and software using simulation to learn models that adapt to each specific task,” says Alberto Rodriguez, an MIT visiting scientist who is a former member of the MechE faculty and now associate director of manipulation research for Boston Dynamics. SimPLE was developed by members of the Manipulation and Mechanisms Lab at MIT (MCube) under Rodriguez’ direction. 

“In this work we show that it is possible to achieve the levels of positional accuracy that are required for many industrial pick and place tasks without any other specialization,” Rodriguez says.

Using a dual-arm robot equipped with visuotactile sensing, the SimPLE solution employs three main components: task-aware grasping, perception by sight and touch (visuotactile perception), and regrasp planning. Real observations are matched against a set of simulated observations through supervised learning so that a distribution of likely object poses can be estimated, and placement accomplished.

In experiments, SimPLE successfully demonstrated the ability to pick-and-place diverse objects spanning a wide range of shapes, achieving successful placements over 90 percent of the time for 6 objects, and over 80 percent of the time for 11 objects.

“There’s an intuitive understanding in the robotics community that vision and touch are both useful, but [until now] there haven’t been many systematic demonstrations of how it can be useful for complex robotics tasks,” says mechanical engineering doctoral student Antonia Delores Bronars SM ’22. Bronars, who is now working with Pulkit Agrawal, assistant professor in the department of Electrical Engineering and Computer Science (EECS), is continuing her PhD work investigating the incorporation of tactile capabilities into robotic systems.

“Most work on grasping ignores the downstream tasks,” says Matt Mason, chief scientist at Berkshire Grey and professor emeritus at Carnegie Mellon University who was not involved in the work. “This paper goes beyond the desire to mimic humans, and shows from a strictly functional viewpoint the utility of combining tactile sensing, vision, with two hands.”

Ken Goldberg, the William S. Floyd Jr. Distinguished Chair in Engineering at the University of California at Berkeley, who was also not involved in the study, says the robot manipulation methodology described in the paper offers a valuable alternative to the trend toward AI and machine learning methods.

“The authors combine well-founded geometric algorithms that can reliably achieve high-precision for a specific set of object shapes and demonstrate that this combination can significantly improve performance over AI methods,” says Goldberg, who is also co-founder and chief scientist for Ambi Robotics and Jacobi Robotics. “This can be immediately useful in industry and is an excellent example of what I call 'good old fashioned engineering' (GOFE).”

Bauza and Bronars say this work was informed by several generations of collaboration.

“In order to really demonstrate how vision and touch can be useful together, it’s necessary to build a full robotic system, which is something that’s very difficult to do as one person over a short horizon of time,” says Bronars. “Collaboration, with each other and with Nikhil [Chavan-Dafle PhD ‘20] and Yifan [Hou PhD ’21 CMU], and across many generations and labs really allowed us to build an end-to-end system.”

The phrase “practice makes perfect” is usually reserved for humans, but it’s also a great maxim for robots newly deployed in unfamiliar environments.

Picture a robot arriving in a warehouse. It comes packaged with the skills it was trained on, like placing an object, and now it needs to pick items from a shelf it’s not familiar with. At first, the machine struggles with this, since it needs to get acquainted with its new surroundings. To improve, the robot will need to understand which skills within an overall task it needs improvement on, then specialize (or parameterize) that action.

A human onsite could program the robot to optimize its performance, but researchers from MIT’s Computer Science and Artificial Intelligence Laboratory (CSAIL) and The AI Institute have developed a more effective alternative. Presented at the Robotics: Science and Systems Conference last month, their “Estimate, Extrapolate, and Situate” (EES) algorithm enables these machines to practice on their own, potentially helping them improve at useful tasks in factories, households, and hospitals. 

Sizing up the situation

To help robots get better at activities like sweeping floors, EES works with a vision system that locates and tracks the machine’s surroundings. Then, the algorithm estimates how reliably the robot executes an action (like sweeping) and whether it would be worthwhile to practice more. EES forecasts how well the robot could perform the overall task if it refines that particular skill, and finally, it practices. The vision system subsequently checks whether that skill was done correctly after each attempt.

EES could come in handy in places like a hospital, factory, house, or coffee shop. For example, if you wanted a robot to clean up your living room, it would need help practicing skills like sweeping. According to Nishanth Kumar SM ’24 and his colleagues, though, EES could help that robot improve without human intervention, using only a few practice trials.

“Going into this project, we wondered if this specialization would be possible in a reasonable amount of samples on a real robot,” says Kumar, co-lead author of a paper describing the work, PhD student in electrical engineering and computer science, and a CSAIL affiliate. “Now, we have an algorithm that enables robots to get meaningfully better at specific skills in a reasonable amount of time with tens or hundreds of data points, an upgrade from the thousands or millions of samples that a standard reinforcement learning algorithm requires.”

See Spot sweep

EES’s knack for efficient learning was evident when implemented on Boston Dynamics’ Spot quadruped during research trials at The AI Institute. The robot, which has an arm attached to its back, completed manipulation tasks after practicing for a few hours. In one demonstration, the robot learned how to securely place a ball and ring on a slanted table in roughly three hours. In another, the algorithm guided the machine to improve at sweeping toys into a bin within about two hours. Both results appear to be an upgrade from previous frameworks, which would have likely taken more than 10 hours per task.

“We aimed to have the robot collect its own experience so it can better choose which strategies will work well in its deployment,” says co-lead author Tom Silver SM ’20, PhD ’24, an electrical engineering and computer science (EECS) alumnus and CSAIL affiliate who is now an assistant professor at Princeton University. “By focusing on what the robot knows, we sought to answer a key question: In the library of skills that the robot has, which is the one that would be most useful to practice right now?”

EES could eventually help streamline autonomous practice for robots in new deployment environments, but for now, it comes with a few limitations. For starters, they used tables that were low to the ground, which made it easier for the robot to see its objects. Kumar and Silver also 3D printed an attachable handle that made the brush easier for Spot to grab. The robot didn’t detect some items and identified objects in the wrong places, so the researchers counted those errors as failures.

Giving robots homework

The researchers note that the practice speeds from the physical experiments could be accelerated further with the help of a simulator. Instead of physically working at each skill autonomously, the robot could eventually combine real and virtual practice. They hope to make their system faster with less latency, engineering EES to overcome the imaging delays the researchers experienced. In the future, they may investigate an algorithm that reasons over sequences of practice attempts instead of planning which skills to refine.

“Enabling robots to learn on their own is both incredibly useful and extremely challenging,” says Danfei Xu, an assistant professor in the School of Interactive Computing at Georgia Tech and a research scientist at NVIDIA AI, who was not involved with this work. “In the future, home robots will be sold to all sorts of households and expected to perform a wide range of tasks. We can't possibly program everything they need to know beforehand, so it’s essential that they can learn on the job. However, letting robots loose to explore and learn without guidance can be very slow and might lead to unintended consequences. The research by Silver and his colleagues introduces an algorithm that allows robots to practice their skills autonomously in a structured way. This is a big step towards creating home robots that can continuously evolve and improve on their own.”

Silver and Kumar’s co-authors are The AI Institute researchers Stephen Proulx and Jennifer Barry, plus four CSAIL members: Northeastern University PhD student and visiting researcher Linfeng Zhao, MIT EECS PhD student Willie McClinton, and MIT EECS professors Leslie Pack Kaelbling and Tomás Lozano-Pérez. Their work was supported, in part, by The AI Institute, the U.S. National Science Foundation, the U.S. Air Force Office of Scientific Research, the U.S. Office of Naval Research, the U.S. Army Research Office, and MIT Quest for Intelligence, with high-performance computing resources from the MIT SuperCloud and Lincoln Laboratory Supercomputing Center.

Many airline passengers naturally worry about flying. But on a worldwide basis, commercial air travel keeps getting safer, according to a new study by MIT researchers. 

The risk of a fatality from commercial air travel was 1 per every 13.7 million passenger boardings globally in the 2018-2022 period — a significant improvement from 1 per 7.9 million boardings in 2008-2017 and a far cry from the 1 per every 350,000 boardings that occurred in 1968-1977, the study finds.

“Aviation safety continues to get better,” says Arnold Barnett, an MIT professor and co-author of a new paper detailing the research results.

“You might think there is some irreducible risk level we can’t get below,” adds Barnett, a leading expert in air travel safety and operations. “And yet, the chance of dying during an air journey keeps dropping by about 7 percent annually, and continues to go down by a factor of two every decade.”

To be sure, there are no guarantees of continual improvement; some recent near-collisions on runways in the U.S. have gained headlines in the last year, making it clear that airline safety is always an ongoing task.

Additionally, the Covid-19 pandemic may have caused a sizable — though presumably temporary — new risk stemming from flying. The study analyzes this risk but quantifies it separately from the long-term safety trend, which is based on accidents and deliberate attacks on aviation.

Overall, Barnett compares these long-run gains in air safety to “Moore’s Law,” the observation that innovators keep finding ways to double the computing power of chips roughly every 18 months. In this case, commercial air travel has gotten roughly twice as safe in each decade dating to the late 1960s.

“Here we have an aerial version of Moore’s Law,” says Barnett, who has helped refine air travel safety statistics for many years.

In per-boarding terms, passengers are about 39 times safer than they were in the 1968-1977 period.

The paper, “Airline safety: Still getting better?” appears in the August issue of the Journal of Air Transport Management. The authors are Barnett, who is the George Eastman Professor of Management Science at the MIT Sloan School of Management, and Jan Reig Torra MBA ’24, a former graduate student at MIT Sloan.

Covid-19 impact

The separate, additional finding about the impact of Covid-19 focuses on cases spread by airline passengers during the pandemic. This is not part of the top-line data, which evaluates airline incidents during normal operations. Still, Barnett thought it would also be valuable to explore the special case of viral transmission during the pandemic.

The study estimates that from June 2020 through February 2021, before vaccines were widely available, there were about 1,200 deaths in the U.S. from Covid-19 associated, directly or indirectly, with its transmission on passenger planes. Most of those fatalities would have involved not passengers but people who got Covid-19 from others who had been infected during air travel.

In addition, the study estimates that from March 2020 through December 2022, around 4,760 deaths around the globe were linked to the transmission of Covid-19 on airplanes. Those estimates are based on the best available data about transmission rates and daily death rates, and take account of the age distributions of air passengers during the pandemic. Perhaps surprisingly, older Americans do not seem to have flown less during the Covid-19 pandemic, even though their risks of death given infection were far higher than those of younger travelers.

“There’s no simple answer to this,” Barnett says. “But we worked to come up with realistic and conservative estimates, so that people can learn important lessons about what happened. I believe people should at least look at these numbers.”

Improved overall safety

Overall, to study fatalities during normal airline operations, the researchers used data from the Flight Safety Foundation, the World Bank, and the International Air Transport Association.

To evaluate air travel risks, experts have used a variety of metrics, including deaths per billion passenger miles, and fatal accidents per 100,000 flight hours. However, Barnett believes deaths per passenger boarding is the most “defensible” and understandable statistic, since it answers a simple question: If you have a boarding pass for a flight, what are your odds of dying? The statistic also includes incidents that might occur in airport terminals.

Having previously developed this metric, Barnett has now updated his findings multiple times, developing a comprehensive picture of air safety over time:

As Barnett’s numbers show, these gains are not incidental improvements, but instead constitute a long-term trend. While the new paper is focused more on empirical outcomes than finding an explanation for them, Barnett suggests there is a combination of factors at work. These include technological advances, such as collision avoidance systems in planes; extensive training; and rigorous work by organizations such as the U.S. Federal Aviation Agency and the National Transportation Safety Board.

However, there are disparities in air travel safety globally. The study divides the world into three tiers of countries, based on their commercial air safety records. For countries in the third tier, there were 36.5 times as many fatalities per passenger boarding in 2018-2022 than was the case in the top tier. Thus, it is safer to fly in some parts of the world than in others.

The first tier of countries consists of the United States, the European Union countries, and other European states, including Montenegro, Norway, Switzerland, and the United Kingdom, as well as Australia, Canada, China, Israel, Japan, and New Zealand.

The second group consists of Bahrain, Bosnia, Brazil, Brunei, Chile, Hong Kong (which has been distinct from mainland China in air safety regulations), India, Jordan, Kuwait, Malaysia, Mexico, the Philippines, Qatar, Singapore, South Africa, South Korea, Taiwan, Thailand, Turkey, and the United Arab Emirates. In each of those two groups of nations, the death risk per boarding over 2018-22 was about 1 per 80 million.

The third group then consists of every other country in the world. Within the top two groups, there were 153 passenger fatalities in the 2018-2022 period, and one major accident, a China Eastern Airlines crash in 2022 that killed 123 passengers. The 30 other fatalities beyond that in the top two tiers stemmed from six other air accidents.

For countries in the third tier, air travel fatalities per boarding were also cut roughly in half during the 2018-2022 period, although, as Barnett noted, that can be interpreted in two ways: It is good they are improving as rapidly as the leading countries in air safety, but in theory, they might be able to apply lessons learned elsewhere and catch up even more quickly.

“The remaining countries continue to improve by something like a factor of two, but they’re still behind the top two groups,” Barnett observes.

Overall, Barnett notes, notwithstanding Covid-19, and looking at accident avoidance, especially in countries with the lowest fatality rates, it is remarkable that air safety keeps getting better. Progress is never assured in this area; yet, the leading countries in air safety, including their government officials and airlines, keep finding ways to make flying safer.

“After decades of sharp improvements, it’s really hard to keep improving at the same rate. And yet they do,” Barnett concludes. 

Electronic waste, or e-waste, is a rapidly growing global problem, and it’s expected to worsen with the production of new kinds of flexible electronics for robotics, wearable devices, health monitors, and other new applications, including single-use devices.

A new kind of flexible substrate material developed at MIT, the University of Utah, and Meta has the potential to enable not only the recycling of materials and components at the end of a device’s useful life, but also the scalable manufacture of more complex multilayered circuits than existing substrates provide.

The development of this new material is described this week in the journal RSC: Applied Polymers, in a paper by MIT Professor Thomas J. Wallin, University of Utah Professor Chen Wang, and seven others.

“We recognize that electronic waste is an ongoing global crisis that’s only going to get worse as we continue to build more devices for the internet of things, and as the rest of the world develops,” says Wallin, an assistant professor in MIT’s Department of Materials Science and Engineering. To date, much academic research on this front has aimed at developing alternatives to conventional substrates for flexible electronics, which primarily use a polymer called Kapton, a trade name for polyimide.

Most such research has focused on entirely different polymer materials, but “that really ignores the commercial side of it, as to why people chose the materials they did to begin with,” Wallin says. Kapton has many advantages, including excellent thermal and insulating properties and ready availability of source materials.

The polyimide business is projected to be a $4 billion global market by 2030. “It’s everywhere, in every electronic device basically,” including parts such as the flexible cables that interconnect different components inside your cellphone or laptop, Wang explains. It’s also widely used in aerospace applications because of its high heat tolerance. “It’s a classic material, but it has not been updated for three or four decades,” he says.

However, it’s also virtually impossible to melt or dissolve Kapton, so it can’t be reprocessed. The same properties also make it harder to manufacture the circuits into advanced architectures, such as multilayered electronics. The traditional way of making Kapton involves heating the material to anywhere from 200 to 300 degrees Celsius. “It’s a rather slow process. It takes hours,” Wang says.

The alternative material that the team developed, which is itself a form of polyimide and therefore should be easily compatible with existing manufacturing infrastructure, is a light-cured polymer similar to those now used by dentists to create tough, durable fillings that cure in a few seconds with ultraviolet light. Not only is this method of hardening the material comparatively fast, it can operate at room temperature.

The new material could serve as the substrate for multilayered circuits, which provides a way of greatly increasing the number of components that can be packed into a small form factor. Previously, since the Kapton substrate doesn’t melt easily, the layers had to be glued together, which adds steps and costs to the process. The fact that the new material can be processed at low-temperature while also hardening very quickly on demand could open up possibilities for new multilayer devices, Wang says.

As for recyclability, the team introduced subunits into the polymer backbone that can be rapidly dissolved away by an alcohol and catalyst solution. Then, precious metals used in the circuits, as well as entire microchips, can be recovered from the solution and reused for new devices.

“We designed the polymer with ester groups in the backbone,” unlike traditional Kapton, Wang explains. These ester groups can be easily broken apart by a fairly mild solution that removes the substrate while leaving the rest of the device unharmed. Wang notes that the University of Utah team has co-founded a company to commercialize the technology. 

“We break the polymer back into its original small molecules. Then we can collect the expensive electronic components and reuse them,” Wallin adds. “We all know about the supply chain shortage with chips and some materials. The rare earth minerals that are in those components are highly valuable. And so we think that there’s a huge economic incentive now, as well as an environmental one, to make these processes for the recapture of these components.”

The research team included Caleb Reese and Grant Musgrave at the University of Utah, and Jenn Wong, Wenyang Pan, John Uehlin, Mason Zadan and Omar Awartani at Meta’s Reality Labs in Redmond, Washington. The work was supported by a startup fund at the Price College of Engineering at the University of Utah.

The MIT School of Science is launching a center to advance knowledge and computational capabilities in the field of sustainability science, and support decision-makers in government, industry, and civil society to achieve sustainable development goals. Aligned with the Climate Project at MIT, researchers at the MIT Center for Sustainability Science and Strategy will develop and apply expertise from across the Institute to improve understanding of sustainability challenges, and thereby provide actionable knowledge and insight to inform strategies for improving human well-being for current and future generations.

Noelle Selin, professor at MIT’s Institute for Data, Systems and Society and the Department of Earth, Atmospheric and Planetary Sciences, will serve as the center’s inaugural faculty director. C. Adam Schlosser and Sergey Paltsev, senior research scientists at MIT, will serve as deputy directors, with Anne Slinn as executive director.

Incorporating and succeeding both the Center for Global Change Science and Joint Program on the Science and Policy of Global Change while adding new capabilities, the center aims to produce leading-edge research to help guide societal transitions toward a more sustainable future. Drawing on the long history of MIT’s efforts to address global change and its integrated environmental and human dimensions, the center is well-positioned to lead burgeoning global efforts to advance the field of sustainability science, which seeks to understand nature-society systems in their full complexity. This understanding is designed to be relevant and actionable for decision-makers in government, industry, and civil society in their efforts to develop viable pathways to improve quality of life for multiple stakeholders.

“As critical challenges such as climate, health, energy, and food security increasingly affect people’s lives around the world, decision-makers need a better understanding of the earth in its full complexity — and that includes people, technologies, and institutions as well as environmental processes,” says Selin. “Better knowledge of these systems and how they interact can lead to more effective strategies that avoid unintended consequences and ensure an improved quality of life for all.”    

Advancing knowledge, computational capability, and decision support

To produce more precise and comprehensive knowledge of sustainability challenges and guide decision-makers to formulate more effective strategies, the center has set the following goals:

• Advance fundamental understanding of the complex interconnected physical and socio-economic systems that affect human well-being. As new policies and technologies are developed amid climate and other global changes, they interact with environmental processes and institutions in ways that can alter the earth’s critical life-support systems. Fundamental mechanisms that determine many of these systems’ behaviors, including those related to interacting climate, water, food, and socio-economic systems, remain largely unknown and poorly quantified. Better understanding can help society mitigate the risks of abrupt changes and “tipping points” in these systems.
• Develop, establish and disseminate new computational tools toward better understanding earth systems, including both environmental and human dimensions. The center’s work will integrate modeling and data analysis across disciplines in an era of increasing volumes of observational data. MIT multi-system models and data products will provide robust information to inform decision-making and shape the next generation of sustainability science and strategy.
• Produce actionable science that supports equity and justice within and across generations. The center’s research will be designed to inform action associated with measurable outcomes aligned with supporting human well-being across generations. This requires engaging a broad range of stakeholders, including not only nations and companies, but also nongovernmental organizations and communities that take action to promote sustainable development — with special attention to those who have historically borne the brunt of environmental injustice.

“The center’s work will advance fundamental understanding in sustainability science, leverage leading-edge computing and data, and promote engagement and impact,” says Selin. “Our researchers will help lead scientists and strategists across the globe who share MIT’s commitment to mobilizing knowledge to inform action toward a more sustainable world.”

Building a better world at MIT

Building on existing MIT capabilities in sustainability science and strategy, the center aims to: 

• focus research, education, and outreach under a theme that reflects a comprehensive state of the field and international research directions, fostering a dynamic community of students, researchers, and faculty;
• raise the visibility of sustainability science at MIT, emphasizing links between science and action, in the context of existing Institute goals and other efforts on climate and sustainability, and in a way that reflects the vital contributions of a range of natural and social science disciplines to understanding human-environment systems; and
• re-emphasize MIT’s long-standing expertise in integrated systems modeling while leveraging the Institute’s concurrent leading-edge strengths in data and computing, establishing leadership that harnesses recent innovations, including those in machine learning and artificial intelligence, toward addressing the science challenges of global change and sustainability.

“The Center for Sustainability Science and Strategy will provide the necessary synergy for our MIT researchers to develop, deploy, and scale up serious solutions to climate change and other critical sustainability challenges,” says Nergis Mavalvala, the Curtis and Kathleen Marble Professor of Astrophysics and dean of the MIT School of Science. “With Professor Selin at its helm, the center will also ensure that these solutions are created in concert with the people who are directly affected now and in the future.”

The center builds on more than three decades of achievements by the Center for Global Change Science and the Joint Program on the Science and Policy of Global Change, both of which were directed or co-directed by professor of atmospheric science Ronald Prinn.

While the moon lacks any breathable air, it does host a barely-there atmosphere. Since the 1980s, astronomers have observed a very thin layer of atoms bouncing over the moon’s surface. This delicate atmosphere — technically known as an “exosphere” — is likely a product of some kind of space weathering. But exactly what those processes might be has been difficult to pin down with any certainty.

Now, scientists at MIT and the University of Chicago say they have identified the main process that formed the moon’s atmosphere and continues to sustain it today. In a study appearing today in Science Advances, the team reports that the lunar atmosphere is primarily a product of “impact vaporization.”

In their study, the researchers analyzed samples of lunar soil collected by astronauts during NASA’s Apollo missions. Their analysis suggests that over the moon’s 4.5-billion-year history its surface has been continuously bombarded, first by massive meteorites, then more recently, by smaller, dust-sized “micrometeoroids.” These constant impacts have kicked up the lunar soil, vaporizing certain atoms on contact and lofting the particles into the air. Some atoms are ejected into space, while others remain suspended over the moon, forming a tenuous atmosphere that is constantly replenished as meteorites continue to pelt the surface.

The researchers found that impact vaporization is the main process by which the moon has generated and sustained its extremely thin atmosphere over billions of years.

“We give a definitive answer that meteorite impact vaporization is the dominant process that creates the lunar atmosphere,” says the study’s lead author, Nicole Nie, an assistant professor in MIT’s Department of Earth, Atmospheric and Planetary Sciences. “The moon is close to 4.5 billion years old, and through that time the surface has been continuously bombarded by meteorites. We show that eventually, a thin atmosphere reaches a steady state because it’s being continuously replenished by small impacts all over the moon.”

Nie’s co-authors are Nicolas Dauphas, Zhe Zhang, and Timo Hopp at the University of Chicago, and Menelaos Sarantos at NASA Goddard Space Flight Center.

Weathering’s roles

In 2013, NASA sent an orbiter around the moon to do some detailed atmospheric reconnaissance. The Lunar Atmosphere and Dust Environment Explorer (LADEE, pronounced “laddie”) was tasked with remotely gathering information about the moon’s thin atmosphere, surface conditions, and any environmental influences on the lunar dust.

LADEE’s mission was designed to determine the origins of the moon’s atmosphere. Scientists hoped that the probe’s remote measurements of soil and atmospheric composition might correlate with certain space weathering processes that could then explain how the moon’s atmosphere came to be.

Researchers suspect that two space weathering processes play a role in shaping the lunar atmosphere: impact vaporization and “ion sputtering” — a phenomenon involving solar wind, which carries energetic charged particles from the sun through space. When these particles hit the moon’s surface, they can transfer their energy to the atoms in the soil and send those atoms sputtering and flying into the air. 

“Based on LADEE’s data, it seemed both processes are playing a role,” Nie says. “For instance, it showed that during meteorite showers, you see more atoms in the atmosphere, meaning impacts have an effect. But it also showed that when the moon is shielded from the sun, such as during an eclipse, there are also changes in the atmosphere’s atoms, meaning the sun also has an impact. So, the results were not clear or quantitative.”

Answers in the soil

To more precisely pin down the lunar atmosphere’s origins, Nie looked to samples of lunar soil collected by astronauts throughout NASA’s Apollo missions. She and her colleagues at the University of Chicago acquired 10 samples of lunar soil, each measuring about 100 milligrams — a tiny amount that she estimates would fit into a single raindrop.

Nie sought to first isolate two elements from each sample: potassium and rubidium. Both elements are “volatile,” meaning that they are easily vaporized by impacts and ion sputtering. Each element exists in the form of several isotopes. An isotope is a variation of the same element, that consists of the same number of protons but a slightly different number of neutrons. For instance, potassium can exist as one of three isotopes, each one having one more neutron, and there being slightly heavier than the last. Similarly, there are two isotopes of rubidium.

The team reasoned that if the moon’s atmosphere consists of atoms that have been vaporized and suspended in the air, lighter isotopes of those atoms should be more easily lofted, while heavier isotopes would be more likely to settle back in the soil. Furthermore, scientists predict that impact vaporization, and ion sputtering, should result in very different isotopic proportions in the soil. The specific ratio of light to heavy isotopes that remain in the soil, for both potassium and rubidium, should then reveal the main process contributing to the lunar atmosphere’s origins.

With all that in mind, Nie analyzed the Apollo samples by first crushing the soils into a fine powder, then dissolving the powders in acids to purify and isolate solutions containing potassium and rubidium. She then passed these solutions through a mass spectrometer to measure the various isotopes of both potassium and rubidium in each sample.

In the end, the team found that the soils contained mostly heavy isotopes of both potassium and rubidium. The researchers were able to quantify the ratio of heavy to light isotopes of both potassium and rubidium, and by comparing both elements, they found that impact vaporization was most likely the dominant process by which atoms are vaporized and lofted to form the moon’s atmosphere.

“With impact vaporization, most of the atoms would stay in the lunar atmosphere, whereas with ion sputtering, a lot of atoms would be ejected into space,” Nie says. “From our study, we now can quantify the role of both processes, to say that the relative contribution of impact vaporization versus ion sputtering is about 70:30 or larger.” In other words, 70 percent or more of the moon’s atmosphere is a product of meteorite impacts, whereas the remaining 30 percent is a consequence of the solar wind.

“The discovery of such a subtle effect is remarkable, thanks to the innovative idea of combining potassium and rubidium isotope measurements along with careful, quantitative modeling,” says Justin Hu, a postdoc who studies lunar soils at Cambridge University, who was not involved in the study. “This discovery goes beyond understanding the moon’s history, as such processes could occur and might be more significant on other moons and asteroids, which are the focus of many planned return missions.”

“Without these Apollo samples, we would not be able to get precise data and measure quantitatively to understand things in more detail,” Nie says. “It’s important for us to bring samples back from the moon and other planetary bodies, so we can draw clearer pictures of the solar system’s formation and evolution.”

This work was supported, in part, by NASA and the National Science Foundation.

Ozone can be an agent of good or harm, depending on where you find it in the atmosphere. Way up in the stratosphere, the colorless gas shields the Earth from the sun’s harsh ultraviolet rays. But closer to the ground, ozone is a harmful air pollutant that can trigger chronic health problems including chest pain, difficulty breathing, and impaired lung function.

And somewhere in between, in the upper troposphere — the layer of the atmosphere just below the stratosphere, where most aircraft cruise — ozone contributes to warming the planet as a potent greenhouse gas.

There are signs that ozone is continuing to rise in the upper troposphere despite efforts to reduce its sources at the surface in many nations. Now, MIT scientists confirm that much of ozone’s increase in the upper troposphere is likely due to humans.

In a paper appearing today in the journal Environmental Science and Technology, the team reports that they detected a clear signal of human influence on upper tropospheric ozone trends in a 17-year satellite record starting in 2005.

“We confirm that there’s a clear and increasing trend in upper tropospheric ozone in the northern midlatitudes due to human beings rather than climate noise,” says study lead author Xinyuan Yu, a graduate student in MIT’s Department of Earth, Atmospheric and Planetary Sciences (EAPS).

“Now we can do more detective work and try to understand what specific human activities are leading to this ozone trend,” adds co-author Arlene Fiore, the Peter H. Stone and Paola Malanotte Stone Professor in Earth, Atmospheric and Planetary Sciences.

The study’s MIT authors include Sebastian Eastham and Qindan Zhu, along with Benjamin Santer at the University of California at Los Angeles, Gustavo Correa of Columbia University, Jean-François Lamarque at the National Center for Atmospheric Research, and Jerald Zimeke at NASA Goddard Space Flight Center.

Ozone’s tangled web

Understanding ozone’s causes and influences is a challenging exercise. Ozone is not emitted directly, but instead is a product of “precursors” — starting ingredients, such as nitrogen oxides and volatile organic compounds (VOCs), that react in the presence of sunlight to form ozone. These precursors are generated from vehicle exhaust, power plants, chemical solvents, industrial processes, aircraft emissions, and other human-induced activities.

Whether and how long ozone lingers in the atmosphere depends on a tangle of variables, including the type and extent of human activities in a given area, as well as natural climate variability. For instance, a strong El Niño year could nudge the atmosphere’s circulation in a way that affects ozone’s concentrations, regardless of how much ozone humans are contributing to the atmosphere that year.

Disentangling the human- versus climate-driven causes of ozone trend, particularly in the upper troposphere, is especially tricky. Complicating matters is the fact that in the lower troposphere — the lowest layer of the atmosphere, closest to ground level — ozone has stopped rising, and has even fallen in some regions at northern midlatitudes in the last few decades. This decrease in lower tropospheric ozone is mainly a result of efforts in North America and Europe to reduce industrial sources of air pollution.

“Near the surface, ozone has been observed to decrease in some regions, and its variations are more closely linked to human emissions,” Yu notes. “In the upper troposphere, the ozone trends are less well-monitored but seem to decouple with those near the surface, and ozone is more easily influenced by climate variability. So, we don’t know whether and how much of that increase in observed ozone in the upper troposphere is attributed to humans.”

A human signal amid climate noise

Yu and Fiore wondered whether a human “fingerprint” in ozone levels, caused directly by human activities, could be strong enough to be detectable in satellite observations in the upper troposphere. To see such a signal, the researchers would first have to know what to look for.

For this, they looked to simulations of the Earth’s climate and atmospheric chemistry. Following approaches developed in climate science, they reasoned that if they could simulate a number of possible climate variations in recent decades, all with identical human-derived sources of ozone precursor emissions, but each starting with a slightly different climate condition, then any differences among these scenarios should be due to climate noise. By inference, any common signal that emerged when averaging over the simulated scenarios should be due to human-driven causes. Such a signal, then, would be a “fingerprint” revealing human-caused ozone, which the team could look for in actual satellite observations.

With this strategy in mind, the team ran simulations using a state-of-the-art chemistry climate model. They ran multiple climate scenarios, each starting from the year 1950 and running through 2014.

From their simulations, the team saw a clear and common signal across scenarios, which they identified as a human fingerprint. They then looked to tropospheric ozone products derived from multiple instruments aboard NASA’s Aura satellite.

“Quite honestly, I thought the satellite data were just going to be too noisy,” Fiore admits. “I didn’t expect that the pattern would be robust enough.”

But the satellite observations they used gave them a good enough shot. The team looked through the upper tropospheric ozone data derived from the satellite products, from the years 2005 to 2021, and found that, indeed, they could see the signal of human-caused ozone that their simulations predicted. The signal is especially pronounced over Asia, where industrial activity has risen significantly in recent decades and where abundant sunlight and frequent weather events loft pollution, including ozone and its precursors, to the upper troposphere.

Yu and Fiore are now looking to identify the specific human activities that are leading to ozone’s increase in the upper troposphere.

“Where is this increasing trend coming from? Is it the near-surface emissions from combusting fossil fuels in vehicle engines and power plants? Is it the aircraft that are flying in the upper troposphere? Is it the influence of wildland fires? Or some combination of all of the above?” Fiore says. “Being able to separate human-caused impacts from natural climate variations can help to inform strategies to address climate change and air pollution.”

This research was funded, in part, by NASA.

MIT physicists and colleagues report new insights into exotic particles key to a form of magnetism that has attracted growing interest because it originates from ultrathin materials only a few atomic layers thick. The work, which could impact future electronics and more, also establishes a new way to study these particles through a powerful instrument at the National Synchrotron Light Source II at Brookhaven National Laboratory.

Among their discoveries, the team has identified the microscopic origin of these particles, known as excitons. They showed how they can be controlled by chemically “tuning” the material, which is primarily composed of nickel. Further, they found that the excitons propagate throughout the bulk material instead of being bound to the nickel atoms.

Finally, they proved that the mechanism behind these discoveries is ubiquitous to similar nickel-based materials, opening the door for identifying — and controlling — new materials with special electronic and magnetic properties.

The open-access results are reported in the July 12 issue of Physical Review X.

“We’ve essentially developed a new research direction into the study of these magnetic two-dimensional materials that very much relies on an advanced spectroscopic method, resonant inelastic X-ray scattering (RIXS), which is available at Brookhaven National Lab,” says Riccardo Comin, MIT’s Class of 1947 Career Development Associate Professor of Physics and leader of the work. Comin is also affiliated with the Materials Research Laboratory and the Research Laboratory of Electronics.

Comin’s colleagues on the work include Connor A. Occhialini, an MIT graduate student in physics, and Yi Tseng, a recent MIT postdoc now at Deutsches Elektronen-Synchrotron (DESY). The two are co-first authors of the Physical Review X paper.

Additional authors are Hebatalla Elnaggar of the Sorbonne; Qian Song, a graduate student in MIT’s Department of Physics; Mark Blei and Seth Ariel Tongay of Arizona State University; Frank M. F. de Groot of Utrecht University; and Valentina Bisogni and Jonathan Pelliciari of Brookhaven National Laboratory.

Ultrathin layers

The magnetic materials at the heart of the current work are known as nickel dihalides. They are composed of layers of nickel atoms sandwiched between layers of halogen atoms (halogens are one family of elements), which can be isolated to atomically thin layers. In this case, the physicists studied the electronic properties of three different materials composed of nickel and the halogens chlorine, bromine, or iodine. Despite their deceptively simple structure, these materials host a rich variety of magnetic phenomena.

The team was interested in how these materials’ magnetic properties respond when exposed to light. They were specifically interested in particular particles — the excitons — and how they are related to the underlying magnetism. How exactly do they form? Can they be controlled?

Enter excitons

A solid material is composed of different types of elementary particles, such as protons and electrons. Also ubiquitous in such materials are “quasiparticles” that the public is less familiar with. These include excitons, which are composed of an electron and a “hole,” or the space left behind when light is shone on a material and energy from a photon causes an electron to jump out of its usual position.

Through the mysteries of quantum mechanics, however, the electron and hole are still connected and can “communicate” with each other through electrostatic interactions. This interaction leads to a new composite particle formed by the electron and the hole — an exciton.

Excitons, unlike electrons, have no charge but possess spin. The spin can be thought of as an elementary magnet, in which the electrons are like little needles orienting in a certain way. In a common refrigerator magnet, the spins all point in the same direction. Generally speaking, the spins can organize in other patterns leading to different kinds of magnets. The unique magnetism associated with the nickel dihalides is one of these less-conventional forms, making it appealing for fundamental and applied research.

The MIT team explored how excitons form in the nickel dihalides. More specifically, they identified the exact energies, or wavelengths, of light necessary for creating them in the three materials they studied.

“We were able to measure and identify the energy necessary to form the excitons in three different nickel halides by chemically ‘tuning,’ or changing, the halide atom from chlorine to bromine to iodine,” says Occhialini. “This is one essential step towards understanding how photons — light — could one day be used to interact with or monitor the magnetic state of these materials.” Ultimate applications include quantum computing and novel sensors.

The work could also help predict new materials involving excitons that might have other interesting properties. Further, while the studied excitons originate on the nickel atoms, the team found that they do not remain localized to these atomic sites. Instead, “we showed that they can effectively hop between sites throughout the crystal,” Occhialini says. “This observation of hopping is the first for these types of excitons, and provides a window into understanding their interplay with the material’s magnetic properties.”

A special instrument

Key to this work — in particular for observing the exciton hopping — is resonant inelastic X-ray scattering (RIXS), an experimental technique that co-authors Pelliciari and Bisogni helped pioneer. Only a few facilities in the world have advanced high energy resolution RIXS instruments. One is at Brookhaven. Pelliciari and Bisogni are part of the team running the RIXS facility at Brookhaven. Occhialini will be joining the team there as a postdoc after receiving his MIT PhD.

RIXS, with its specific sensitivity to the excitons from the nickel atoms, allowed the team to “set the basis for a general framework for nickel dihalide systems,” says Pelliciari. “it allowed us to directly measure the propagation of excitons.”

This work was supported by the U.S. Department of Energy Basic Energy Science and Brookhaven National Laboratory through the Co-design Center for Quantum Advantage (C2QA), a DoE Quantum Information Science Research Center.

Shimmering ice extends in all directions as far as the eye can see. Air temperatures plunge to minus 40 degrees Fahrenheit and colder with wind chills. Ocean currents drag large swaths of ice floating at sea. Polar bears, narwhals, and other iconic Arctic species roam wild.

For a week this past spring, MIT Lincoln Laboratory researchers Ben Evans and Dave Whelihan called this place — drifting some 200 nautical miles offshore from Prudhoe Bay, Alaska, on the frozen Beaufort Sea in the Arctic Circle — home. Two ice runways for small aircraft provided their only way in and out of this remote wilderness; heated tents provided their only shelter from the bitter cold.

Here, in the northernmost region on Earth, Evans and Whelihan joined other groups conducting fieldwork in the Arctic as part of Operation Ice Camp (OIC) 2024, an operational exercise run by the U.S. Navy's Arctic Submarine Laboratory (ASL). Riding on snowmobiles and helicopters, the duo deployed a small set of integrated sensor nodes that measure everything from atmospheric conditions to ice properties to the structure of water deep below the surface.

Ultimately, they envision deploying an unattended network of these low-cost sensor nodes across the Arctic to increase scientific understanding of the trending loss in sea ice extent and thickness. Warming much faster than the rest of the world, the Arctic is a ground zero for climate change, with cascading impacts across the planet that include rising sea levels and extreme weather. Openings in the sea ice cover, or leads, are concerning not only for climate change but also for global geopolitical competition over transit routes and natural resources. A synoptic view of the physical processes happening above, at, and below sea ice is key to determining why the ice is diminishing. In turn, this knowledge can help predict when and where fractures will occur, to inform planning and decision-making.

Winter “camp”

Every two years, OIC, previously called Ice Exercise (ICEX), provides a way for the international community to access the Arctic for operational readiness exercises and scientific research, with the focus switching back and forth; this year’s focus was scientific research. Coordination, planning, and execution of the month-long operation is led by ASL, a division of the U.S. Navy’s Undersea Warfighting Development Center responsible for ensuring the submarine force can effectively operate in the Arctic Ocean.

Making this inhospitable and unforgiving environment safe for participants takes considerable effort. The critical first step is determining where to set up camp. In the weeks before the first participants arrived for OIC 2024, ASL — with assistance from the U.S. National Ice Center, University of Alaska Fairbanks Geophysical Institute, and UIC Science — flew over large sheets of floating ice (ice floes) identified via satellite imagery, landed on some they thought might be viable sites, and drilled through the ice to check its thickness. The ice floe must not only be large enough to accommodate construction of a camp and two runways but also feature both multiyear ice and first-year ice. Multiyear ice is thick and strong but rough, making it ideal for camp setup, while the smooth but thinner first-year ice is better suited for building runways. Once the appropriate ice floe was selected, ASL began to haul in equipment and food, build infrastructure like lodging and a command center, and fly in a small group before fully operationalizing the site. They also identified locations near the camp for two Navy submarines to surface through the ice.

The more than 200 participants represented U.S. and allied forces and scientists from research organizations and universities. Distinguished visitors from government offices also attended OIC to see the unique Arctic environment and unfolding challenges firsthand.

“Our ASL hosts do incredible work to build this camp from scratch and keep us alive,” Evans says.

Evans and Whelihan, part of the laboratory’s Advanced Undersea Systems and Technology Group, first trekked to the Arctic in March 2022 for ICEX 2022. (The laboratory in general has been participating since 2016 in these events, the first iteration of which occurred in 1946.) There, they deployed a suite of commercial off-the-shelf sensors for detecting acoustic (sound) and seismic (vibration) events created by ice fractures or collisions, and for measuring salinity, temperature, and pressure in the water below the ice. They also deployed a prototype fiber-based temperature sensor array developed by the laboratory and research partners for precisely measuring temperature across the entire water column at one location, and a University of New Hampshire (UNH)−supplied echosounder to investigate the different layers present in the water column. In this maiden voyage, their goals were to assess how these sensors fared in the harsh Arctic conditions and to collect a dataset from which characteristic signatures of ice-fracturing events could begin to be identified. These events would be correlated with weather and water conditions to eventually offer a predictive capability.

“We saw real phenomenology in our data,” Whelihan says. “But, we’re not ice experts. What we’re good at here at the laboratory is making and deploying sensors. That's our place in the world of climate science: to be a data provider. In fact, we hope to open source all of our data this year so that ice scientists can access and analyze them and then we can make enhanced sensors and collect more data.”

Interim ice

In the two years since that expedition, they and their colleagues have been modifying their sensor designs and deployment strategies. As Evans and Whelihan learned at ICEX 2022, to be resilient in the Arctic, a sensor must not only be kept warm and dry during deployment but also be deployed in a way to prevent breaking. Moreover, sufficient power and data links are needed to collect and access sensor data.

“We can make cold-weather electronics, no problem,” Whelihan says. “The two drivers are operating the sensors in an energy-starved environment — the colder it is, the worse batteries perform — and keeping them from getting destroyed when ice floes crash together as leads in the ice open up.”

Their work in the interim to OIC 2024 involved integrating the individual sensors into hardened sensor nodes and practicing deploying these nodes in easier-to-access locations. To facilitate incorporating additional sensors into a node, Whelihan spearheaded the development of an open-source, easily extensible hardware and software architecture.

In March 2023, the Lincoln Laboratory team deployed three sensor nodes for a week on Huron Bay off Lake Superior through Michigan Tech's Great Lakes Research Center (GLRC). Engineers from GLRC helped the team safely set up an operations base on the ice. They demonstrated that the sensor integration worked, and the sensor nodes proved capable of surviving for at least a week in relatively harsh conditions. The researchers recorded seismic activity on all three nodes, corresponding to some ice breaking further up the bay.

“Proving our sensor node in an Arctic surrogate environment provided a stepping stone for testing in the real Arctic,” Evans says.

Evans then received an invitation from Ignatius Rigor, the coordinator of the International Arctic Buoy Program (IABP), to join him on an upcoming trip to Utqiaġvik (formerly Barrow), Alaska, and deploy one of their seismic sensor nodes on the ice there (with support from UIC Science). The IABP maintains a network of Arctic buoys equipped with meteorological and oceanic sensors. Data collected by these buoys are shared with the operational and research communities to support real-time operations (e.g., forecasting sea ice conditions for coastal Alaskans) and climate research. However, these buoys are typically limited in the frequency at which they collect data, so phenomenology on shorter time scales important to climate change may be missed. Moreover, these buoys are difficult and expensive to deploy because they are designed to survive in the harshest environments for years at a time.  

The laboratory-developed sensor nodes could offer an inexpensive, easier-to-deploy option for collecting more data over shorter periods of time. In April 2023, Evans placed a sensor node in Utqiaġvik on landfast sea ice, which is stationary ice anchored to the seabed just off the coast. During the sensor node’s week-long deployment, a big piece of drift ice (ice not attached to the seabed or other fixed object) broke off and crashed into the landfast ice. The event was recorded by a radar maintained by the University of Alaska Fairbanks that monitors sea ice movement in near real time to warn of any instability. Though this phenomenology is not exactly the same as that expected for Arctic sea ice, the researchers were encouraged to see seismic activity recorded by their sensor node.

In December 2023, Evans and Whelihan headed to New Hampshire, where they conducted echosounder testing in UNH’s engineering test tank and on the Piscataqua River. Together with their UNH partners, they sought to determine whether a low-cost, hobby-grade echosounder could detect the same phenomenology of interest as the high-fidelity UNH echosounder, which would be far too costly to deploy in sensor nodes across the Arctic. In the test tank and on the river, the low-cost echosounder proved capable of detecting masses of water moving in the water column, but with considerably less structural detail than afforded by the higher-cost option. Seeing such dynamics is important to inferring where water comes from and understanding how it affects sea ice breakup — for example, how warm water moving in from the Pacific Ocean is coming into contact with and melting the ice. So, the laboratory researchers and UNH partners have been building a medium-fidelity, medium-cost echosounder.

In January 2024, Evans and Whelihan — along with Jehan Diaz, a fellow staff member in their research group — returned to GLRC. With logistical support from their GLRC hosts, they snowmobiled across the ice on Portage Lake, where they practiced several activities to prepare for OIC 2024: augering (drilling) six-inch holes in the ice, albeit in thinner ice than that in the Arctic; placing their long, pipe-like sensor nodes through these holes; operating cold-hardened drones to interact with the nodes; and retrieving the nodes. They also practiced sensor calibration by hitting the ice with an iron bar some distance away from the nodes and correlating this distance with the resulting measured acoustic and seismic intensity.

“Our time at GLRC helped us mitigate a lot of risks and prepare to deploy these complex systems in the Arctic,” Whelihan says.

Arctic again

To get to OIC, Evans and Whelihan first flew to Prudhoe Bay and reacclimated to the frigid temperatures. They spent the next two days at the Deadhorse Aviation Center hangar inspecting their equipment for transit-induced damage, which included squashed cables and connectors that required rejiggering.

“That’s part of the adventure story,” Evans says. “Getting stuff to Prudhoe Bay is not your standard shipping; it’s ice-road trucking.”

From there, they boarded a small aircraft to the ice camp.

“Even though this trip marked our second time coming here, it was still disorienting,” Evans continues. "You land in the middle of nowhere on a small aircraft after a couple-hour flight. You get out bundled in all of your Arctic gear in this remote, pristine environment.”

After unloading and rechecking their equipment for any damage, calibrating their sensors, and attending safety briefings, they were ready to begin their experiments.

An icy situation

Inside the project tent, Evans and Whelihan deployed the UNH-supplied echosounder and a suite of ground-truth sensors on an automated winch to profile water conductivity, temperature, and depth (CTD). Echosounder data needed to be validated with associated CTD data to determine the source of the water in the water column. Ocean properties change as a function of depth, and these changes are important to capture, in part because masses of water coming in from the Atlantic and Pacific oceans arrive at different depths. Though masses of warm water have always existed, climate change–related mechanisms are now bringing them into contact with the ice.  

“As ice breaks up, wind can directly interact with the ocean because it’s lacking that barrier of ice cover,” Evans explains. “Kinetic energy from the wind causes mixing in the ocean; all the warm water that used to stay at depth instead gets brought up and interacts with the ice.”

They also deployed four of their sensor nodes several miles outside of camp. To access this deployment site, they rode on a sled pulled via a snowmobile driven by Ann Hill, an ASL field party leader trained in Arctic survival and wildlife encounters. The temperature that day was -55 F. At such a dangerously cold temperature, frostnip and frostbite are all too common. To avoid removal of gloves or other protective clothing, the researchers enabled the nodes with WiFi capability (the nodes also have a satellite communications link to transmit low-bandwidth data). Large amounts of data are automatically downloaded over WiFi to an arm-wearable haptic (touch-based) system when a user walks up to a node.

“It was so cold that the holes we were drilling in the ice to reach the water column were freezing solid,” Evans explains. “We realized it was going to be quite an ordeal to get our sensor nodes out of the ice.”

So, after drilling a big hole in the ice, they deployed only one central node with all the sensor components: a commercial echosounder, an underwater microphone, a seismometer, and a weather station. They deployed the other three nodes, each with a seismometer and weather station, atop the ice.

“One of our design considerations was flexibility,” Whelihan says. “Each node can integrate as few or as many sensors as desired.”

The small sensor array was only collecting data for about a day when Evans and Whelihan, who were at the time on a helicopter, saw that their initial field site had become completely cut off from camp by a 150-meter-wide ice lead. They quickly returned to camp to load the tools needed to pull the nodes, which were no longer accessible by snowmobile. Two recently arrived staff members from the Ted Stevens Center for Arctic Security Studies offered to help them retrieve their nodes. The helicopter landed on the ice floe near a crack, and the pilot told them they had half an hour to complete their recovery mission. By the time they had retrieved all four sensors, the crack had increased from thumb to fist size.

“When we got home, we analyzed the collected sensor data and saw a spike in seismic activity corresponding to what could be the major ice-fracturing event that necessitated our node recovery mission,” Whelihan says.  

The researchers also conducted experiments with their Arctic-hardened drones to evaluate their utility for retrieving sensor node data and to develop concepts of operations for future capabilities.

“The idea is to have some autonomous vehicle land next to the node, download data, and come back, like a data mule, rather than having to expend energy getting data off the system, say via high-speed satellite communications,” Whelihan says. “We also started testing whether the drone is capable on its own of finding sensors that are constantly moving and getting close enough to them. Even flying in 25-mile-per-hour winds, and at very low temperatures, the drone worked well.”

Aside from carrying out their experiments, the researchers had the opportunity to interact with other participants. Their “roommates” were ice scientists from Norway and Finland. They met other ice and water scientists conducting chemistry experiments on the salt content of ice taken from different depths in the ice sheet (when ocean water freezes, salt tends to get pushed out of the ice). One of their collaborators — Nicholas Schmerr, an ice seismologist from the University of Maryland — placed high-quality geophones (for measuring vibrations in the ice) alongside their nodes deployed on the camp field site. They also met with junior enlisted submariners, who temporarily came to camp to open up spots on the submarine for distinguished visitors.

“Part of what we've been doing over the last three years is building connections within the Arctic community,” Evans says. “Every time I start to get a handle on the phenomenology that exists out here, I learn something new. For example, I didn’t know that sometimes a layer of ice forms a little bit deeper than the primary ice sheet, and you can actually see fish swimming in between the layers.”

“One day, we were out with our field party leader, who saw fog while she was looking at the horizon and said the ice was breaking up,” Whelihan adds. “I said, 'Wait, what?' As she explained, when an ice lead forms, fog comes out of the ocean. Sure enough, within 30 minutes, we had quarter-mile visibility, whereas beforehand it was unlimited.”

Back to solid ground

Before leaving, Whelihan and Evans retrieved and packed up all the remaining sensor nodes, adopting the “leave no trace” philosophy of preserving natural places.

“Only a limited number of people get access to this special environment,” Whelihan says. “We hope to grow our footprint at these events in future years, giving opportunities to other laboratory staff members to attend.”

In the meantime, they will analyze the collected sensor data and refine their sensor node design. One design consideration is how to replenish the sensors’ battery power. A potential path forward is to leverage the temperature difference between water and air, and harvest energy from the water currents moving under ice floes. Wind energy may provide another viable solution. Solar power would only work for part of the year because the Arctic Circle undergoes periods of complete darkness.

The team is also seeking external sponsorship to continue their work engineering sensing systems that advance the scientific community’s understanding of changes to Arctic ice; this work is currently funded through Lincoln Laboratory's internally administered R&D portfolio on climate change. And, in learning more about this changing environment and its critical importance to strategic interests, they are considering other sensing problems that they could tackle using their Arctic engineering expertise.

“The Arctic is becoming a more visible and important region because of how it’s changing,” Evans concludes. “Going forward as a country, we must be able to operate there.”

People use large language models for a huge array of tasks, from translating an article to identifying financial fraud. However, despite the incredible capabilities and versatility of these models, they sometimes generate inaccurate responses.

On top of that problem, the models can be overconfident about wrong answers or underconfident about correct ones, making it tough for a user to know when a model can be trusted.

Researchers typically calibrate a machine-learning model to ensure its level of confidence lines up with its accuracy. A well-calibrated model should have less confidence about an incorrect prediction, and vice-versa. But because large language models (LLMs) can be applied to a seemingly endless collection of diverse tasks, traditional calibration methods are ineffective.

Now, researchers from MIT and the MIT-IBM Watson AI Lab have introduced a calibration method tailored to large language models. Their method, called Thermometer, involves building a smaller, auxiliary model that runs on top of a large language model to calibrate it.

Thermometer is more efficient than other approaches — requiring less power-hungry computation — while preserving the accuracy of the model and enabling it to produce better-calibrated responses on tasks it has not seen before.

By enabling efficient calibration of an LLM for a variety of tasks, Thermometer could help users pinpoint situations where a model is overconfident about false predictions, ultimately preventing them from deploying that model in a situation where it may fail.

“With Thermometer, we want to provide the user with a clear signal to tell them whether a model’s response is accurate or inaccurate, in a way that reflects the model’s uncertainty, so they know if that model is reliable,” says Maohao Shen, an electrical engineering and computer science (EECS) graduate student and lead author of a paper on Thermometer.

Shen is joined on the paper by Gregory Wornell, the Sumitomo Professor of Engineering who leads the Signals, Information, and Algorithms Laboratory in the Research Laboratory for Electronics, and is a member of the MIT-IBM Watson AI Lab; senior author Soumya Ghosh, a research staff member in the MIT-IBM Watson AI Lab; as well as others at MIT and the MIT-IBM Watson AI Lab. The research was recently presented at the International Conference on Machine Learning.

Universal calibration

Since traditional machine-learning models are typically designed to perform a single task, calibrating them usually involves one task-specific method. On the other hand, since LLMs have the flexibility to perform many tasks, using a traditional method to calibrate that model for one task might hurt its performance on another task.

Calibrating an LLM often involves sampling from the model multiple times to obtain different predictions and then aggregating these predictions to obtain better-calibrated confidence. However, because these models have billions of parameters, the computational costs of such approaches rapidly add up.

“In a sense, large language models are universal because they can handle various tasks. So, we need a universal calibration method that can also handle many different tasks,” says Shen.

With Thermometer, the researchers developed a versatile technique that leverages a classical calibration method called temperature scaling to efficiently calibrate an LLM for a new task.

In this context, a “temperature” is a scaling parameter used to adjust a model’s confidence to be aligned with its prediction accuracy. Traditionally, one determines the right temperature using a labeled validation dataset of task-specific examples.

Since LLMs are often applied to new tasks, labeled datasets can be nearly impossible to acquire. For instance, a user who wants to deploy an LLM to answer customer questions about a new product likely does not have a dataset containing such questions and answers.

Instead of using a labeled dataset, the researchers train an auxiliary model that runs on top of an LLM to automatically predict the temperature needed to calibrate it for this new task.

They use labeled datasets of a few representative tasks to train the Thermometer model, but then once it has been trained, it can generalize to new tasks in a similar category without the need for additional labeled data.

A Thermometer model trained on a collection of multiple-choice question datasets, perhaps including one with algebra questions and one with medical questions, could be used to calibrate an LLM that will answer questions about geometry or biology, for instance.

“The aspirational goal is for it to work on any task, but we are not quite there yet,” Ghosh says.   

The Thermometer model only needs to access a small part of the LLM’s inner workings to predict the right temperature that will calibrate its prediction for data points of a specific task. 

An efficient approach

Importantly, the technique does not require multiple training runs and only slightly slows the LLM. Plus, since temperature scaling does not alter a model’s predictions, Thermometer preserves its accuracy.

When they compared Thermometer to several baselines on multiple tasks, it consistently produced better-calibrated uncertainty measures while requiring much less computation.

“As long as we train a Thermometer model on a sufficiently large number of tasks, it should be able to generalize well across any new task, just like a large language model, it is also a universal model,” Shen adds.

The researchers also found that if they train a Thermometer model for a smaller LLM, it can be directly applied to calibrate a larger LLM within the same family.

In the future, they want to adapt Thermometer for more complex text-generation tasks and apply the technique to even larger LLMs. The researchers also hope to quantify the diversity and number of labeled datasets one would need to train a Thermometer model so it can generalize to a new task.

This research was funded, in part, by the MIT-IBM Watson AI Lab.

If patients receiving intensive care or undergoing major surgery develop excessively high or low blood pressures, they could suffer severe organ dysfunction. It’s not enough for their care team to know that pressure is abnormal. To choose the correct drug to treat the problem, doctors must know why blood pressure has changed. A new MIT study presents the mathematical framework needed to derive that crucial information accurately and in real time.

The mathematical approach, described in a recent open-access study in IEEE Transactions on Biomedical Engineering, produces proportional estimates of the two critical factors underlying blood pressure changes: the heart’s rate of blood output (cardiac output) and the arterial system’s resistance to that blood flow (systemic vascular resistance). By applying the new method to previously collected data from animal models, the researchers show that their estimates, derived from minimally invasive measures of peripheral arterial blood pressure, accurately matched estimates using additional information from an invasive flow probe placed on the aorta. Moreover, the estimates accurately tracked the changes induced in the animals by the various drugs physicians use to correct aberrant blood pressure.

“Estimates of resistance and cardiac output from our approach provide information that can readily be used to guide hemodynamic management decisions in real time,” the study authors wrote.

With further testing leading to regulatory approval, the authors say, the method would be applicable during heart surgeries, liver transplants, intensive care unit treatment, and many other procedures affecting cardiovascular function or blood volume.

“Any patient who is having cardiac surgery could need this,” says study senior author Emery N. Brown, the Edward Hood Taplin Professor of Medical Engineering and Computational Neuroscience in The Picower Institute for Learning and Memory, the Institute for Medical Engineering and Science, and the Department of Brain and Cognitive Sciences at MIT. Brown is also an anesthesiologist at Massachusetts General Hospital and a professor of anesthesiology at Harvard Medical School. “So might any patient undergoing a more normal surgery but who might have a compromised cardiovascular system, such as ischemic heart disease. You can’t have the blood pressure being all over the place.”

The study’s lead author is electrical engineering and computer science (EECS) graduate student Taylor Baum, who is co-supervised by Brown and Munther Dahleh, the William A. Coolidge Professor in EECS.

Algorithmic advance

The idea that cardiac output and systemic resistance are the two key components of blood pressure comes from the two-element Windkessel model. The new study is not the first to use the model to estimate these components from blood pressure measurements, but previous attempts ran into a trade-off between quick estimate updates and the accuracy of estimates; methods would either provide more erroneous estimates at every beat or more reliable estimates that are updated at minute time scales. Led by Baum, the MIT team overcame the trade-off with a new approach of applying statistical and signal processing techniques such as “state-space” modeling.

“Our estimates, updated at every beat, are not just informed by the current beat; but they incorporate where things were in previous beats as well,” Baum says. “It’s that combination of past history and current observations that produces a more reliable estimate while still at a beat-by-beat time scale.”

Notably, the resulting estimates of cardiac output and systemic resistance are “proportional,” meaning that they are each inextricably linked in the math with another co-factor, rather than estimated on their own. But application of the new method to data collected in an older study from six animals showed that the proportional estimates from recordings using minimally invasive catheters provide comparable information for cardiovascular system management.

One key finding was that the proportional estimates made based on arterial blood pressure readings from catheters inserted in various locations away from the heart (e.g., the leg or the arm) mirrored estimates derived from more invasive catheters placed within the aorta. The significance of the finding is that a system using the new estimation method could in some cases rely on a minimally invasive catheter in various peripheral arteries, thereby avoiding the need for a riskier placement of a central artery catheter or a pulmonary artery catheter directly in the heart, the clinical gold standard for cardiovascular state estimation.

Another key finding was that when the animals received each of five drugs that doctors use to regulate either systemic vascular resistance or cardiac output, the proportional estimates tracked the resulting changes properly. The finding therefore suggests that the proportional estimates of each factor are accurately reflecting their physiological changes.

Toward the clinic

With these encouraging results, Baum and Brown say, the current method can be readily implemented in clinical settings to inform perioperative care teams about underlying causes of critical blood pressure changes. They are actively pursuing regulatory approval of use of this method in a clinical device.

Additionally, the researchers are pursuing more animal studies to validate an advanced blood pressure management approach that uses this method. They have developed a closed-loop system, informed by this estimation framework, to precisely regulate blood pressure in an animal model. Upon completion of the animal studies, they will apply for regulatory clearance to test the system in humans.

In addition to Baum, Dahleh and Brown, the paper’s other authors are Elie Adam, Christian Guay, Gabriel Schamberg, Mohammadreza Kazemi, and Thomas Heldt.

The National Science Foundation, the National Institutes of Health, a Mathworks Fellowship, The Picower Institute for Learning and Memory, and The JPB Foundation supported the study.

In 2021, a team led by MIT physicists reported creating a new ultrathin ferroelectric material, or one where positive and negative charges separate into different layers. At the time they noted the material’s potential for applications in computer memory and much more. Now the same core team and colleagues — including two from the lab next door — have built a transistor with that material and shown that its properties are so useful that it could change the world of electronics.

Although the team’s results are based on a single transistor in the lab, “in several aspects its properties already meet or exceed industry standards” for the ferroelectric transistors produced today, says Pablo Jarillo-Herrero, the Cecil and Ida Green Professor of Physics, who led the work with professor of physics Raymond Ashoori. Both are also affiliated with the Materials Research Laboratory.

“In my lab we primarily do fundamental physics. This is one of the first, and perhaps most dramatic, examples of how very basic science has led to something that could have a major impact on applications,” Jarillo-Herrero says.

Says Ashoori, “When I think of my whole career in physics, this is the work that I think 10 to 20 years from now could change the world.”

Among the new transistor’s superlative properties:

• It can switch between positive and negative charges — essentially the ones and zeros of digital information — at very high speeds, on nanosecond time scales. (A nanosecond is a billionth of a second.)
• It is extremely tough. After 100 billion switches it still worked with no signs of degradation.
• The material behind the magic is only billionths of a meter thick, one of the thinnest of its kind in the world. That, in turn, could allow for much denser computer memory storage. It could also lead to much more energy-efficient transistors because the voltage required for switching scales with material thickness. (Ultrathin equals ultralow voltages.)

The work is reported in a recent issue of Science. The co-first authors of the paper are Kenji Yasuda, now an assistant professor at Cornell University, and Evan Zalys-Geller, now at Atom Computing. Additional authors are Xirui Wang, an MIT graduate student in physics; Daniel Bennett and Efthimios Kaxiras of Harvard University; Suraj S. Cheema, an assistant professor in MIT’s Department of Electrical Engineering and Computer Science and an affiliate of the Research Laboratory of Electronics; and Kenji Watanabe and Takashi Taniguchi of the National Institute for Materials Science in Japan.

What they did

In a ferroelectric material, positive and negative charges spontaneously head to different sides, or poles. Upon the application of an external electric field, those charges switch sides, reversing the polarization. Switching the polarization can be used to encode digital information, and that information will be nonvolatile, or stable over time. It won’t change unless an electric field is applied. For a ferroelectric to have broad application to electronics, all of this needs to happen at room temperature.

The new ferroelectric material reported in Science in 2021 is based on atomically thin sheets of boron nitride that are stacked parallel to each other, a configuration that doesn’t exist in nature. In bulk boron nitride, the individual layers of boron nitride are instead rotated by 180 degrees.

It turns out that when an electric field is applied to this parallel stacked configuration, one layer of the new boron nitride material slides over the other, slightly changing the positions of the boron and nitrogen atoms. For example, imagine that each of your hands is composed of only one layer of cells. The new phenomenon is akin to pressing your hands together then slightly shifting one above the other.

“So the miracle is that by sliding the two layers a few angstroms, you end up with radically different electronics,” says Ashoori. The diameter of an atom is about 1 angstrom.

Another miracle: “nothing wears out in the sliding,” Ashoori continues. That’s why the new transistor could be switched 100 billion times without degrading. Compare that to the memory in a flash drive made with conventional materials. “Each time you write and erase a flash memory, you get some degradation,” says Ashoori. “Over time, it wears out, which means that you have to use some very sophisticated methods for distributing where you’re reading and writing on the chip.” The new material could make those steps obsolete.

A collaborative effort

Yasuda, the co-first author of the current Science paper, applauds the collaborations involved in the work. Among them, “we [Jarillo-Herrero’s team] made the material and, together with Ray [Ashoori] and [co-first author] Evan [Zalys-Geller], we measured its characteristics in detail. That was very exciting.” Says Ashoori, “many of the techniques in my lab just naturally applied to work that was going on in the lab next door. It’s been a lot of fun.”

Ashoori notes that “there’s a lot of interesting physics behind this” that could be explored. For example, “if you think about the two layers sliding past each other, where does that sliding start?” In addition, says Yasuda, could the ferroelectricity be triggered with something other than electricity, like an optical pulse? And is there a fundamental limit to the amount of switches the material can make?

Challenges remain. For example, the current way of producing the new ferroelectrics is difficult and not conducive to mass manufacturing. “We made a single transistor as a demonstration. If people could grow these materials on the wafer scale, we could create many, many more,” says Yasuda. He notes that different groups are already working to that end.

Concludes Ashoori, “There are a few problems. But if you solve them, this material fits in so many ways into potential future electronics. It’s very exciting.”

This work was supported by the U.S. Army Research Office, the MIT/Microsystems Technology Laboratories Samsung Semiconductor Research Fund, the U.S. National Science Foundation, the Gordon and Betty Moore Foundation, the Ramon Areces Foundation, the Basic Energy Sciences program of the U.S. Department of Energy, the Japan Society for the Promotion of Science, and the Ministry of Education, Culture, Sports, Science and Technology (MEXT) of Japan.

A sustainable source for clean energy may lie in old soda cans and seawater.

MIT engineers have found that when the aluminum in soda cans is exposed in its pure form and mixed with seawater, the solution bubbles up and naturally produces hydrogen — a gas that can be subsequently used to power an engine or fuel cell without generating carbon emissions. What’s more, this simple reaction can be sped up by adding a common stimulant: caffeine.

In a study appearing today in the journal Cell Reports Physical Science, the researchers show they can produce hydrogen gas by dropping pretreated, pebble-sized aluminum pellets into a beaker of filtered seawater. The aluminum is pretreated with a rare-metal alloy that effectively scrubs aluminum into a pure form that can react with seawater to generate hydrogen. The salt ions in the seawater can in turn attract and recover the alloy, which can be reused to generate more hydrogen, in a sustainable cycle.

The team found that this reaction between aluminum and seawater successfully produces hydrogen gas, though slowly. On a lark, they tossed into the mix some coffee grounds and found, to their surprise, that the reaction picked up its pace.

In the end, the team discovered that a low concentration of imidazole — an active ingredient in caffeine — is enough to significantly speed up the reaction, producing the same amount of hydrogen in just five minutes, compared to two hours without the added stimulant.

The researchers are developing a small reactor that could run on a marine vessel or underwater vehicle. The vessel would hold a supply of aluminum pellets (recycled from old soda cans and other aluminum products), along with a small amount of gallium-indium and caffeine. These ingredients could be periodically funneled into the reactor, along with some of the surrounding seawater, to produce hydrogen on demand. The hydrogen could then fuel an onboard engine to drive a motor or generate electricity to power the ship.

“This is very interesting for maritime applications like boats or underwater vehicles because you wouldn’t have to carry around seawater — it’s readily available,” says study lead author Aly Kombargi, a PhD student in MIT’s Department of Mechanical Engineering. “We also don’t have to carry a tank of hydrogen. Instead, we would transport aluminum as the ‘fuel,’ and just add water to produce the hydrogen that we need.”

The study’s co-authors include Enoch Ellis, an undergraduate in chemical engineering; Peter Godart PhD ’21, who has founded a company to recycle aluminum as a source of hydrogen fuel; and Douglas Hart, MIT professor of mechanical engineering.

Shields up

The MIT team, led by Hart, is developing efficient and sustainable methods to produce hydrogen gas, which is seen as a “green” energy source that could power engines and fuel cells without generating climate-warming emissions.

One drawback to fueling vehicles with hydrogen is that some designs would require the gas to be carried onboard like traditional gasoline in a tank — a risky setup, given hydrogen’s volatile potential. Hart and his team have instead looked for ways to power vehicles with hydrogen without having to constantly transport the gas itself.

They found a possible workaround in aluminum — a naturally abundant and stable material that, when in contact with water, undergoes a straightforward chemical reaction that generates hydrogen and heat.

The reaction, however, comes with a sort of Catch-22: While aluminum can generate hydrogen when it mixes with water, it can only do so in a pure, exposed state. The instant aluminum meets with oxygen, such as in air, the surface immediately forms a thin, shield-like layer of oxide that prevents further reactions. This barrier is the reason hydrogen doesn’t immediately bubble up when you drop a soda can in water.

In previous work, using fresh water, the team found they could pierce aluminum’s shield and keep the reaction with water going by pretreating the aluminum with a small amount of rare metal alloy made from a specific concentration of gallium and indium. The alloy serves as an “activator,” scrubbing away any oxide buildup and creating a pure aluminum surface that is free to react with water. When they ran the reaction in fresh, de-ionized water, they found that one pretreated pellet of aluminum produced 400 milliliters of hydrogen in just five minutes. They estimate that just 1 gram of pellets would generate 1.3 liters of hydrogen in the same amount of time.

But to further scale up the system would require a significant supply of gallium indium, which is relatively expensive and rare.

“For this idea to be cost-effective and sustainable, we had to work on recovering this alloy postreaction,” Kombargi says.

By the sea

In the team’s new work, they found they could retrieve and reuse gallium indium using a solution of ions. The ions — atoms or molecules with an electrical charge — protect the metal alloy from reacting with water and help it to precipitate into a form that can be scooped out and reused.   

“Lucky for us, seawater is an ionic solution that is very cheap and available,” says Kombargi, who tested the idea with seawater from a nearby beach. “I literally went to Revere Beach with a friend and we grabbed our bottles and filled them, and then I just filtered out algae and sand, added aluminum to it, and it worked with the same consistent results.”

He found that hydrogen indeed bubbled up when he added aluminum to a beaker of filtered seawater. And he was able to scoop out the gallium indium afterward. But the reaction happened much more slowly than it did in fresh water. It turns out that the ions in seawater act to shield gallium indium, such that it can coalesce and be recovered after the reaction. But the ions have a similar effect on aluminum, building up a barrier that slows its reaction with water.

As they looked for ways to speed up the reaction in seawater, the researchers tried out various and unconventional ingredients.

“We were just playing around with things in the kitchen, and found that when we added coffee grounds into seawater and dropped aluminum pellets in, the reaction was quite fast compared to just seawater,” Kombargi says.

To see what might explain the speedup, the team reached out to colleagues in MIT’s chemistry department, who suggested they try imidazole — an active ingredient in caffeine, which happens to have a molecular structure that can pierce through aluminum (allowing the material to continue reacting with water), while leaving gallium indium’s ionic shield intact.

“That was our big win,” Kombargi says. “We had everything we wanted: recovering the gallium indium, plus the fast and efficient reaction.”

The researchers believe they have the essential ingredients to run a sustainable hydrogen reactor. They plan to test it first in marine and underwater vehicles. They’ve calculated that such a reactor, holding about 40 pounds of aluminum pellets, could power a small underwater glider for about 30 days by pumping in surrounding seawater and generating hydrogen to power a motor.

“We’re showing a new way to produce hydrogen fuel, without carrying hydrogen but carrying aluminum as the ‘fuel,’” Kombargi says. “The next part is to figure out how to use this for trucks, trains, and maybe airplanes. Perhaps, instead of having to carry water as well, we could extract water from the ambient humidity to produce hydrogen. That’s down the line.”

An open-access MIT study published today in Nature provides new evidence for how specific cells and circuits become vulnerable in Alzheimer’s disease, and hones in on other factors that may help some people show resilience to cognitive decline, even amid clear signs of disease pathology. 

To highlight potential targets for interventions to sustain cognition and memory, the authors engaged in a novel comparison of gene expression across multiple brain regions in people with or without Alzheimer’s disease, and conducted lab experiments to test and validate their major findings.

Brain cells all have the same DNA but what makes them differ, both in their identity and their activity, are their patterns of how they express those genes. The new analysis measured gene expression differences in more than 1.3 million cells of more than 70 cell types in six brain regions from 48 tissue donors, 26 of whom died with an Alzheimer’s diagnosis and 22 of whom without. As such, the study provides a uniquely large, far-ranging, and yet detailed accounting of how brain cell activity differs amid Alzheimer’s disease by cell type, by brain region, by disease pathology, and by each person’s cognitive assessment while still alive.

“Specific brain regions are vulnerable in Alzheimer’s and there is an important need to understand how these regions or particular cell types are vulnerable,” says co-senior author Li-Huei Tsai, Picower Professor of Neuroscience and director of The Picower Institute for Learning and Memory and the Aging Brain Initiative at MIT. “And the brain is not just neurons. It’s many other cell types. How these cell types may respond differently, depending on where they are, is something fascinating we are only at the beginning of looking at.”

Co-senior author Manolis Kellis, professor of computer science and head of MIT’s Computational Biology Group, likens the technique used to measure gene expression comparisons, single-cell RNA profiling, to being a much more advanced “microscope” than the ones that first allowed Alois Alzheimer to characterize the disease’s pathology more than a century ago.

“Where Alzheimer saw amyloid protein plaques and phosphorylated tau tangles in his microscope, our single-cell ‘microscope’ tells us, cell by cell and gene by gene, about thousands of subtle yet important biological changes in response to pathology,” says Kellis. “Connecting this information with the cognitive state of patients reveals how cellular responses relate with cognitive loss or resilience, and can help propose new ways to treat cognitive loss. Pathology can precede cognitive symptoms by a decade or two before cognitive decline becomes diagnosed. If there’s not much we can do about the pathology at that stage, we can at least try to safeguard the cellular pathways that maintain cognitive function.”

Hansruedi Mathys, a former MIT postdoc in the Tsai Lab who is now an assistant professor at the University of Pittsburgh; Carles Boix PhD '22, a former graduate student in Kellis’s lab who is now a postdoc at Harvard Medical School; and Leyla Akay, a graduate student in Tsai’s lab, led the study analyzing the prefrontal cortex, entorhinal cortex, hippocampus, anterior thalamus, angular gyrus, and the midtemporal cortex. The brain samples came from the Religious Order Study and the Rush Memory and Aging Project at Rush University.

Neural vulnerability and Reelin

Some of the earliest signs of amyloid pathology and neuron loss in Alzheimer’s occur in memory-focused regions called the hippocampus and the entorhinal cortex. In those regions, and in other parts of the cerebral cortex, the researchers were able to pinpoint a potential reason why. One type of excitatory neuron in the hippocampus and four in the entorhinal cortex were significantly less abundant in people with Alzheimer’s than in people without. Individuals with depletion of those cells performed significantly worse on cognitive assessments. Moreover, many vulnerable neurons were interconnected in a common neuronal circuit. And just as importantly, several either directly expressed a protein called Reelin, or were directly affected by Reelin signaling. In all, therefore, the findings distinctly highlight especially vulnerable neurons, whose loss is associated with reduced cognition, that share a neuronal circuit and a molecular pathway.

Tsai notes that Reelin has become prominent in Alzheimer’s research because of a recent study of a man in Colombia. He had a rare mutation in the Reelin gene that caused the protein to be more active, and was able to stay cognitively healthy at an advanced age despite having a strong family predisposition to early-onset Alzheimer’s. The new study shows that loss of Reelin-producing neurons is associated with cognitive decline. Taken together, it might mean that the brain benefits from Reelin, but that neurons that produce it may be lost in at least some Alzheimer’s patients.

“We can think of Reelin as having maybe some kind of protective or beneficial effect,” Akay says. “But we don’t yet know what it does or how it could confer resilience.”

In further analysis the researchers also found that specifically vulnerable inhibitory neuron subtypes identified in a previously study from this group in the prefrontal cortex also were involved in Reelin signaling, further reinforcing the significance of the molecule and its signaling pathway.

To further check their results, the team directly examined the human brain tissue samples and the brains of two kinds of Alzheimer’s model mice. Sure enough, those experiments also showed a reduction in Reelin-positive neurons in the human and mouse entorhinal cortex.

Resilience associated with choline metabolism in astrocytes

To find factors that might preserve cognition, even amid pathology, the team examined which genes, in which cells, and in which regions, were most closely associated with cognitive resilience, which they defined as residual cognitive function, above the typical cognitive loss expected given the observed pathology.

Their analysis yielded a surprising and specific answer: across several brain regions, astrocytes that expressed genes associated with antioxidant activity and with choline metabolism and polyamine biosynthesis were significantly associated with sustained cognition, even amid high levels of tau and amyloid. The results reinforced previous research findings led by Tsai and Susan Lundqvist in which they showed that dietary supplement of choline helped astrocytes cope with the dysregulation of lipids caused by the most significant Alzheimer’s risk gene, the APOE4 variant. The antioxidant findings also pointed to a molecule that can be found as a dietary supplement, spermidine, which may have anti-inflammatory properties, although such an association would need further work to be established causally.

As before, the team went beyond the predictions from the single-cell RNA expression analysis to make direct observations in the brain tissue of samples. Those that came from cognitively resilient individuals indeed showed increased expression of several of the astrocyte-expressed genes predicted to be associated with cognitive resilience.

New analysis method, open dataset

To analyze the mountains of single-cell data, the researchers developed a new robust methodology based on groups of coordinately-expressed genes (known as “gene modules”), thus exploiting the expression correlation patterns between functionally-related genes in the same module.

“In principle, the 1.3 million cells we surveyed could use their 20,000 genes in an astronomical number of different combinations,” explains Kellis. “In practice, however, we observe a much smaller subset of coordinated changes. Recognizing these coordinated patterns allow us to infer much more robust changes, because they are based on multiple genes in the same functionally-connected module.”

He offered this analogy: With many joints in their bodies, people could move in all kinds of crazy ways, but in practice they engage in many fewer coordinated movements like walking, running, or dancing. The new method enables scientists to identify such coordinated gene expression programs as a group.

While Kellis and Tsai’s labs already reported several noteworthy findings from the dataset, the researchers expect that many more possibly significant discoveries still wait to be found in the trove of data. To facilitate such discovery the team posted handy analytical and visualization tools along with the data on Kellis’s website.

“The dataset is so immensely rich. We focused on only a few aspects that are salient that we believe are very, very interesting, but by no means have we exhausted what can be learned with this dataset,” Kellis says. “We expect many more discoveries ahead, and we hope that young researchers (of all ages) will dive right in and surprise us with many more insights.”

Going forward, Kellis says, the researchers are studying the control circuitry associated with the differentially expressed genes, to understand the genetic variants, the regulators, and other driver factors that can be modulated to reverse disease circuitry across brain regions, cell types, and different stages of the disease.

Additional authors of the study include Ziting Xia, Jose Davila Velderrain, Ayesha P. Ng, Xueqiao Jiang, Ghada Abdelhady, Kyriaki Galani, Julio Mantero, Neil Band, Benjamin T. James, Sudhagar Babu, Fabiola Galiana-Melendez, Kate Louderback, Dmitry Prokopenko, Rudolph E. Tanzi, and David A. Bennett.

Support for the research came from the National Institutes of Health, The Picower Institute for Learning and Memory, The JPB Foundation, the Cure Alzheimer’s Fund, The Robert A. and Renee E. Belfer Family Foundation, Eduardo Eurnekian, and Joseph DiSabato.

Organizations are increasingly utilizing machine-learning models to allocate scarce resources or opportunities. For instance, such models can help companies screen resumes to choose job interview candidates or aid hospitals in ranking kidney transplant patients based on their likelihood of survival.

When deploying a model, users typically strive to ensure its predictions are fair by reducing bias. This often involves techniques like adjusting the features a model uses to make decisions or calibrating the scores it generates.

However, researchers from MIT and Northeastern University argue that these fairness methods are not sufficient to address structural injustices and inherent uncertainties. In a new paper, they show how randomizing a model’s decisions in a structured way can improve fairness in certain situations.

For example, if multiple companies use the same machine-learning model to rank job interview candidates deterministically — without any randomization — then one deserving individual could be the bottom-ranked candidate for every job, perhaps due to how the model weighs answers provided in an online form. Introducing randomization into a model’s decisions could prevent one worthy person or group from always being denied a scarce resource, like a job interview.

Through their analysis, the researchers found that randomization can be especially beneficial when a model’s decisions involve uncertainty or when the same group consistently receives negative decisions.

They present a framework one could use to introduce a specific amount of randomization into a model’s decisions by allocating resources through a weighted lottery. This method, which an individual can tailor to fit their situation, can improve fairness without hurting the efficiency or accuracy of a model.

“Even if you could make fair predictions, should you be deciding these social allocations of scarce resources or opportunities strictly off scores or rankings? As things scale, and we see more and more opportunities being decided by these algorithms, the inherent uncertainties in these scores can be amplified. We show that fairness may require some sort of randomization,” says Shomik Jain, a graduate student in the Institute for Data, Systems, and Society (IDSS) and lead author of the paper.

Jain is joined on the paper by Kathleen Creel, assistant professor of philosophy and computer science at Northeastern University; and senior author Ashia Wilson, the Lister Brothers Career Development Professor in the Department of Electrical Engineering and Computer Science and a principal investigator in the Laboratory for Information and Decision Systems (LIDS). The research will be presented at the International Conference on Machine Learning.

Considering claims

This work builds off a previous paper in which the researchers explored harms that can occur when one uses deterministic systems at scale. They found that using a machine-learning model to deterministically allocate resources can amplify inequalities that exist in training data, which can reinforce bias and systemic inequality. 

“Randomization is a very useful concept in statistics, and to our delight, satisfies the fairness demands coming from both a systemic and individual point of view,” Wilson says.

In this paper, they explored the question of when randomization can improve fairness. They framed their analysis around the ideas of philosopher John Broome, who wrote about the value of using lotteries to award scarce resources in a way that honors all claims of individuals.

A person’s claim to a scarce resource, like a kidney transplant, can stem from merit, deservingness, or need. For instance, everyone has a right to life, and their claims on a kidney transplant may stem from that right, Wilson explains.

“When you acknowledge that people have different claims to these scarce resources, fairness is going to require that we respect all claims of individuals. If we always give someone with a stronger claim the resource, is that fair?” Jain says.

That sort of deterministic allocation could cause systemic exclusion or exacerbate patterned inequality, which occurs when receiving one allocation increases an individual’s likelihood of receiving future allocations. In addition, machine-learning models can make mistakes, and a deterministic approach could cause the same mistake to be repeated.

Randomization can overcome these problems, but that doesn’t mean all decisions a model makes should be randomized equally.

Structured randomization

The researchers use a weighted lottery to adjust the level of randomization based on the amount of uncertainty involved in the model’s decision-making. A decision that is less certain should incorporate more randomization.

“In kidney allocation, usually the planning is around projected lifespan, and that is deeply uncertain. If two patients are only five years apart, it becomes a lot harder to measure. We want to leverage that level of uncertainty to tailor the randomization,” Wilson says.

The researchers used statistical uncertainty quantification methods to determine how much randomization is needed in different situations. They show that calibrated randomization can lead to fairer outcomes for individuals without significantly affecting the utility, or effectiveness, of the model.

“There is a balance to be had between overall utility and respecting the rights of the individuals who are receiving a scarce resource, but oftentimes the tradeoff is relatively small,” says Wilson.

However, the researchers emphasize there are situations where randomizing decisions would not improve fairness and could harm individuals, such as in criminal justice contexts.

But there could be other areas where randomization can improve fairness, such as college admissions, and the researchers plan to study other use cases in future work. They also want to explore how randomization can affect other factors, such as competition or prices, and how it could be used to improve the robustness of machine-learning models.

“We are hoping our paper is a first move toward illustrating that there might be a benefit to randomization. We are offering randomization as a tool. How much you are going to want to do it is going to be up to all the stakeholders in the allocation to decide. And, of course, how they decide is another research question all together,” says Wilson.

As artificial intelligence models become increasingly prevalent and are integrated into diverse sectors like health care, finance, education, transportation, and entertainment, understanding how they work under the hood is critical. Interpreting the mechanisms underlying AI models enables us to audit them for safety and biases, with the potential to deepen our understanding of the science behind intelligence itself.

Imagine if we could directly investigate the human brain by manipulating each of its individual neurons to examine their roles in perceiving a particular object. While such an experiment would be prohibitively invasive in the human brain, it is more feasible in another type of neural network: one that is artificial. However, somewhat similar to the human brain, artificial models containing millions of neurons are too large and complex to study by hand, making interpretability at scale a very challenging task. 

To address this, MIT Computer Science and Artificial Intelligence Laboratory (CSAIL) researchers decided to take an automated approach to interpreting artificial vision models that evaluate different properties of images. They developed “MAIA” (Multimodal Automated Interpretability Agent), a system that automates a variety of neural network interpretability tasks using a vision-language model backbone equipped with tools for experimenting on other AI systems.

“Our goal is to create an AI researcher that can conduct interpretability experiments autonomously. Existing automated interpretability methods merely label or visualize data in a one-shot process. On the other hand, MAIA can generate hypotheses, design experiments to test them, and refine its understanding through iterative analysis,” says Tamar Rott Shaham, an MIT electrical engineering and computer science (EECS) postdoc at CSAIL and co-author on a new paper about the research. “By combining a pre-trained vision-language model with a library of interpretability tools, our multimodal method can respond to user queries by composing and running targeted experiments on specific models, continuously refining its approach until it can provide a comprehensive answer.”

The automated agent is demonstrated to tackle three key tasks: It labels individual components inside vision models and describes the visual concepts that activate them, it cleans up image classifiers by removing irrelevant features to make them more robust to new situations, and it hunts for hidden biases in AI systems to help uncover potential fairness issues in their outputs. “But a key advantage of a system like MAIA is its flexibility,” says Sarah Schwettmann PhD ’21, a research scientist at CSAIL and co-lead of the research. “We demonstrated MAIA’s usefulness on a few specific tasks, but given that the system is built from a foundation model with broad reasoning capabilities, it can answer many different types of interpretability queries from users, and design experiments on the fly to investigate them.” 

Neuron by neuron

In one example task, a human user asks MAIA to describe the concepts that a particular neuron inside a vision model is responsible for detecting. To investigate this question, MAIA first uses a tool that retrieves “dataset exemplars” from the ImageNet dataset, which maximally activate the neuron. For this example neuron, those images show people in formal attire, and closeups of their chins and necks. MAIA makes various hypotheses for what drives the neuron’s activity: facial expressions, chins, or neckties. MAIA then uses its tools to design experiments to test each hypothesis individually by generating and editing synthetic images — in one experiment, adding a bow tie to an image of a human face increases the neuron’s response. “This approach allows us to determine the specific cause of the neuron’s activity, much like a real scientific experiment,” says Rott Shaham.

MAIA’s explanations of neuron behaviors are evaluated in two key ways. First, synthetic systems with known ground-truth behaviors are used to assess the accuracy of MAIA’s interpretations. Second, for “real” neurons inside trained AI systems with no ground-truth descriptions, the authors design a new automated evaluation protocol that measures how well MAIA’s descriptions predict neuron behavior on unseen data.

The CSAIL-led method outperformed baseline methods describing individual neurons in a variety of vision models such as ResNet, CLIP, and the vision transformer DINO. MAIA also performed well on the new dataset of synthetic neurons with known ground-truth descriptions. For both the real and synthetic systems, the descriptions were often on par with descriptions written by human experts.

How are descriptions of AI system components, like individual neurons, useful? “Understanding and localizing behaviors inside large AI systems is a key part of auditing these systems for safety before they’re deployed — in some of our experiments, we show how MAIA can be used to find neurons with unwanted behaviors and remove these behaviors from a model,” says Schwettmann. “We’re building toward a more resilient AI ecosystem where tools for understanding and monitoring AI systems keep pace with system scaling, enabling us to investigate and hopefully understand unforeseen challenges introduced by new models.”

Peeking inside neural networks

The nascent field of interpretability is maturing into a distinct research area alongside the rise of “black box” machine learning models. How can researchers crack open these models and understand how they work?

Current methods for peeking inside tend to be limited either in scale or in the precision of the explanations they can produce. Moreover, existing methods tend to fit a particular model and a specific task. This caused the researchers to ask: How can we build a generic system to help users answer interpretability questions about AI models while combining the flexibility of human experimentation with the scalability of automated techniques?

One critical area they wanted this system to address was bias. To determine whether image classifiers displayed bias against particular subcategories of images, the team looked at the final layer of the classification stream (in a system designed to sort or label items, much like a machine that identifies whether a photo is of a dog, cat, or bird) and the probability scores of input images (confidence levels that the machine assigns to its guesses). To understand potential biases in image classification, MAIA was asked to find a subset of images in specific classes (for example “labrador retriever”) that were likely to be incorrectly labeled by the system. In this example, MAIA found that images of black labradors were likely to be misclassified, suggesting a bias in the model toward yellow-furred retrievers.

Since MAIA relies on external tools to design experiments, its performance is limited by the quality of those tools. But, as the quality of tools like image synthesis models improve, so will MAIA. MAIA also shows confirmation bias at times, where it sometimes incorrectly confirms its initial hypothesis. To mitigate this, the researchers built an image-to-text tool, which uses a different instance of the language model to summarize experimental results. Another failure mode is overfitting to a particular experiment, where the model sometimes makes premature conclusions based on minimal evidence.

“I think a natural next step for our lab is to move beyond artificial systems and apply similar experiments to human perception,” says Rott Shaham. “Testing this has traditionally required manually designing and testing stimuli, which is labor-intensive. With our agent, we can scale up this process, designing and testing numerous stimuli simultaneously. This might also allow us to compare human visual perception with artificial systems.”

“Understanding neural networks is difficult for humans because they have hundreds of thousands of neurons, each with complex behavior patterns. MAIA helps to bridge this by developing AI agents that can automatically analyze these neurons and report distilled findings back to humans in a digestible way,” says Jacob Steinhardt, assistant professor at the University of California at Berkeley, who wasn’t involved in the research. “Scaling these methods up could be one of the most important routes to understanding and safely overseeing AI systems.”

Rott Shaham and Schwettmann are joined by five fellow CSAIL affiliates on the paper: undergraduate student Franklin Wang; incoming MIT student Achyuta Rajaram; EECS PhD student Evan Hernandez SM ’22; and EECS professors Jacob Andreas and Antonio Torralba. Their work was supported, in part, by the MIT-IBM Watson AI Lab, Open Philanthropy, Hyundai Motor Co., the Army Research Laboratory, Intel, the National Science Foundation, the Zuckerman STEM Leadership Program, and the Viterbi Fellowship. The researchers’ findings will be presented at the International Conference on Machine Learning this week.

As the name suggests, most electronic devices today work through the movement of electrons. But materials that can efficiently conduct protons — the nucleus of the hydrogen atom — could be key to a number of important technologies for combating global climate change.

Most proton-conducting inorganic materials available now require undesirably high temperatures to achieve sufficiently high conductivity. However, lower-temperature alternatives could enable a variety of technologies, such as more efficient and durable fuel cells to produce clean electricity from hydrogen, electrolyzers to make clean fuels such as hydrogen for transportation, solid-state proton batteries, and even new kinds of computing devices based on iono-electronic effects.

In order to advance the development of proton conductors, MIT engineers have identified certain traits of materials that give rise to fast proton conduction. Using those traits quantitatively, the team identified a half-dozen new candidates that show promise as fast proton conductors. Simulations suggest these candidates will perform far better than existing materials, although they still need to be conformed experimentally. In addition to uncovering potential new materials, the research also provides a deeper understanding at the atomic level of how such materials work.

The new findings are described in the journal Energy and Environmental Sciences, in a paper by MIT professors Bilge Yildiz and Ju Li, postdocs Pjotrs Zguns and Konstantin Klyukin, and their collaborator Sossina Haile and her students from Northwestern University. Yildiz is the Breene M. Kerr Professor in the departments of Nuclear Science and Engineering, and Materials Science and Engineering.

“Proton conductors are needed in clean energy conversion applications such as fuel cells, where we use hydrogen to produce carbon dioxide-free electricity,” Yildiz explains. “We want to do this process efficiently, and therefore we need materials that can transport protons very fast through such devices.”

Present methods of producing hydrogen, for example steam methane reforming, emit a great deal of carbon dioxide. “One way to eliminate that is to electrochemically produce hydrogen from water vapor, and that needs very good proton conductors,” Yildiz says. Production of other important industrial chemicals and potential fuels, such as ammonia, can also be carried out through efficient electrochemical systems that require good proton conductors.

But most inorganic materials that conduct protons can only operate at temperatures of 200 to 600 degrees Celsius (roughly 450 to 1,100 Fahrenheit), or even higher. Such temperatures require energy to maintain and can cause degradation of materials. “Going to higher temperatures is not desirable because that makes the whole system more challenging, and the material durability becomes an issue,” Yildiz says. “There is no good inorganic proton conductor at room temperature.” Today, the only known room-temperature proton conductor is a polymeric material that is not practical for applications in computing devices because it can’t easily be scaled down to the nanometer regime, she says.

To tackle the problem, the team first needed to develop a basic and quantitative understanding of exactly how proton conduction works, taking a class of inorganic proton conductors, called solid acids. “One has to first understand what governs proton conduction in these inorganic compounds,” she says. While looking at the materials’ atomic configurations, the researchers identified a pair of characteristics that directly relates to the materials’ proton-carrying potential.

As Yildiz explains, proton conduction first involves a proton “hopping from a donor oxygen atom to an acceptor oxygen. And then the environment has to reorganize and take the accepted proton away, so that it can hop to another neighboring acceptor, enabling long-range proton diffusion.” This process happens in many inorganic solids, she says. Figuring out how that last part works — how the atomic lattice gets reorganized to take the accepted proton away from the original donor atom — was a key part of this research, she says.

The researchers used computer simulations to study a class of materials called solid acids that become good proton conductors above 200 degrees Celsius. This class of materials has a substructure called the polyanion group sublattice, and these groups have to rotate and take the proton away from its original site so it can then transfer to other sites. The researchers were able to identify the phonons that contribute to the flexibility of this sublattice, which is essential for proton conduction. Then they used this information to comb through vast databases of theoretically and experimentally possible compounds, in search of better proton conducting materials.

As a result, they found solid acid compounds that are promising proton conductors and that have been developed and produced for a variety of different applications but never before studied as proton conductors; these compounds turned out to have just the right characteristics of lattice flexibility. The team then carried out computer simulations of how the specific materials they identified in their initial screening would perform under relevant temperatures, to confirm their suitability as proton conductors for fuel cells or other uses. Sure enough, they found six promising materials, with predicted proton conduction speeds faster than the best existing solid acid proton conductors.

“There are uncertainties in these simulations,” Yildiz cautions. “I don’t want to say exactly how much higher the conductivity will be, but these look very promising. Hopefully this motivates the experimental field to try to synthesize them in different forms and make use of these compounds as proton conductors.”

Translating these theoretical findings into practical devices could take some years, she says. The likely first applications would be for electrochemical cells to produce fuels and chemical feedstocks such as hydrogen and ammonia, she says.

The work was supported by the U.S. Department of Energy, the Wallenberg Foundation, and the U.S. National Science Foundation.

One thing that makes large language models (LLMs) so powerful is the diversity of tasks to which they can be applied. The same machine-learning model that can help a graduate student draft an email could also aid a clinician in diagnosing cancer.

However, the wide applicability of these models also makes them challenging to evaluate in a systematic way. It would be impossible to create a benchmark dataset to test a model on every type of question it can be asked.

In a new paper, MIT researchers took a different approach. They argue that, because humans decide when to deploy large language models, evaluating a model requires an understanding of how people form beliefs about its capabilities.

For example, the graduate student must decide whether the model could be helpful in drafting a particular email, and the clinician must determine which cases would be best to consult the model on.

Building off this idea, the researchers created a framework to evaluate an LLM based on its alignment with a human’s beliefs about how it will perform on a certain task.

They introduce a human generalization function — a model of how people update their beliefs about an LLM’s capabilities after interacting with it. Then, they evaluate how aligned LLMs are with this human generalization function.

Their results indicate that when models are misaligned with the human generalization function, a user could be overconfident or underconfident about where to deploy it, which might cause the model to fail unexpectedly. Furthermore, due to this misalignment, more capable models tend to perform worse than smaller models in high-stakes situations.

“These tools are exciting because they are general-purpose, but because they are general-purpose, they will be collaborating with people, so we have to take the human in the loop into account,” says study co-author Ashesh Rambachan, assistant professor of economics and a principal investigator in the Laboratory for Information and Decision Systems (LIDS).

Rambachan is joined on the paper by lead author Keyon Vafa, a postdoc at Harvard University; and Sendhil Mullainathan, an MIT professor in the departments of Electrical Engineering and Computer Science and of Economics, and a member of LIDS. The research will be presented at the International Conference on Machine Learning.

Human generalization

As we interact with other people, we form beliefs about what we think they do and do not know. For instance, if your friend is finicky about correcting people’s grammar, you might generalize and think they would also excel at sentence construction, even though you’ve never asked them questions about sentence construction.

“Language models often seem so human. We wanted to illustrate that this force of human generalization is also present in how people form beliefs about language models,” Rambachan says.

As a starting point, the researchers formally defined the human generalization function, which involves asking questions, observing how a person or LLM responds, and then making inferences about how that person or model would respond to related questions.

If someone sees that an LLM can correctly answer questions about matrix inversion, they might also assume it can ace questions about simple arithmetic. A model that is misaligned with this function — one that doesn’t perform well on questions a human expects it to answer correctly — could fail when deployed.

With that formal definition in hand, the researchers designed a survey to measure how people generalize when they interact with LLMs and other people.

They showed survey participants questions that a person or LLM got right or wrong and then asked if they thought that person or LLM would answer a related question correctly. Through the survey, they generated a dataset of nearly 19,000 examples of how humans generalize about LLM performance across 79 diverse tasks.

Measuring misalignment

They found that participants did quite well when asked whether a human who got one question right would answer a related question right, but they were much worse at generalizing about the performance of LLMs.

“Human generalization gets applied to language models, but that breaks down because these language models don’t actually show patterns of expertise like people would,” Rambachan says.

People were also more likely to update their beliefs about an LLM when it answered questions incorrectly than when it got questions right. They also tended to believe that LLM performance on simple questions would have little bearing on its performance on more complex questions.

In situations where people put more weight on incorrect responses, simpler models outperformed very large models like GPT-4.

“Language models that get better can almost trick people into thinking they will perform well on related questions when, in actuality, they don’t,” he says.

One possible explanation for why humans are worse at generalizing for LLMs could come from their novelty — people have far less experience interacting with LLMs than with other people.

“Moving forward, it is possible that we may get better just by virtue of interacting with language models more,” he says.

To this end, the researchers want to conduct additional studies of how people’s beliefs about LLMs evolve over time as they interact with a model. They also want to explore how human generalization could be incorporated into the development of LLMs.

“When we are training these algorithms in the first place, or trying to update them with human feedback, we need to account for the human generalization function in how we think about measuring performance,” he says.

In the meanwhile, the researchers hope their dataset could be used a benchmark to compare how LLMs perform related to the human generalization function, which could help improve the performance of models deployed in real-world situations.

“To me, the contribution of the paper is twofold. The first is practical: The paper uncovers a critical issue with deploying LLMs for general consumer use. If people don’t have the right understanding of when LLMs will be accurate and when they will fail, then they will be more likely to see mistakes and perhaps be discouraged from further use. This highlights the issue of aligning the models with people's understanding of generalization,” says Alex Imas, professor of behavioral science and economics at the University of Chicago’s Booth School of Business, who was not involved with this work. “The second contribution is more fundamental: The lack of generalization to expected problems and domains helps in getting a better picture of what the models are doing when they get a problem ‘correct.’ It provides a test of whether LLMs ‘understand’ the problem they are solving.”

This research was funded, in part, by the Harvard Data Science Initiative and the Center for Applied AI at the University of Chicago Booth School of Business.

Ductal carcinoma in situ (DCIS) is a type of preinvasive tumor that sometimes progresses to a highly deadly form of breast cancer. It accounts for about 25 percent of all breast cancer diagnoses.

Because it is difficult for clinicians to determine the type and stage of DCIS, patients with DCIS are often overtreated. To address this, an interdisciplinary team of researchers from MIT and ETH Zurich developed an AI model that can identify the different stages of DCIS from a cheap and easy-to-obtain breast tissue image. Their model shows that both the state and arrangement of cells in a tissue sample are important for determining the stage of DCIS.

Because such tissue images are so easy to obtain, the researchers were able to build one of the largest datasets of its kind, which they used to train and test their model. When they compared its predictions to conclusions of a pathologist, they found clear agreement in many instances.

In the future, the model could be used as a tool to help clinicians streamline the diagnosis of simpler cases without the need for labor-intensive tests, giving them more time to evaluate cases where it is less clear if DCIS will become invasive.

“We took the first step in understanding that we should be looking at the spatial organization of cells when diagnosing DCIS, and now we have developed a technique that is scalable. From here, we really need a prospective study. Working with a hospital and getting this all the way to the clinic will be an important step forward,” says Caroline Uhler, a professor in the Department of Electrical Engineering and Computer Science (EECS) and the Institute for Data, Systems, and Society (IDSS), who is also director of the Eric and Wendy Schmidt Center at the Broad Institute of MIT and Harvard and a researcher at MIT’s Laboratory for Information and Decision Systems (LIDS).

Uhler, co-corresponding author of a paper on this research, is joined by lead author Xinyi Zhang, a graduate student in EECS and the Eric and Wendy Schmidt Center; co-corresponding author GV Shivashankar, professor of mechogenomics at ETH Zurich jointly with the Paul Scherrer Institute; and others at MIT, ETH Zurich, and the University of Palermo in Italy. The open-access research was published July 20 in Nature Communications.

Combining imaging with AI   

Between 30 and 50 percent of patients with DCIS develop a highly invasive stage of cancer, but researchers don’t know the biomarkers that could tell a clinician which tumors will progress.

Researchers can use techniques like multiplexed staining or single-cell RNA sequencing to determine the stage of DCIS in tissue samples. However, these tests are too expensive to be performed widely, Shivashankar explains.

In previous work, these researchers showed that a cheap imagining technique known as chromatin staining could be as informative as the much costlier single-cell RNA sequencing.

For this research, they hypothesized that combining this single stain with a carefully designed machine-learning model could provide the same information about cancer stage as costlier techniques.

First, they created a dataset containing 560 tissue sample images from 122 patients at three different stages of disease. They used this dataset to train an AI model that learns a representation of the state of each cell in a tissue sample image, which it uses to infer the stage of a patient’s cancer.

However, not every cell is indicative of cancer, so the researchers had to aggregate them in a meaningful way.

They designed the model to create clusters of cells in similar states, identifying eight states that are important markers of DCIS. Some cell states are more indicative of invasive cancer than others. The model determines the proportion of cells in each state in a tissue sample.

Organization matters

“But in cancer, the organization of cells also changes. We found that just having the proportions of cells in every state is not enough. You also need to understand how the cells are organized,” says Shivashankar.

With this insight, they designed the model to consider proportion and arrangement of cell states, which significantly boosted its accuracy.

“The interesting thing for us was seeing how much spatial organization matters. Previous studies had shown that cells which are close to the breast duct are important. But it is also important to consider which cells are close to which other cells,” says Zhang.

When they compared the results of their model with samples evaluated by a pathologist, it had clear agreement in many instances. In cases that were not as clear-cut, the model could provide information about features in a tissue sample, like the organization of cells, that a pathologist could use in decision-making.

This versatile model could also be adapted for use in other types of cancer, or even neurodegenerative conditions, which is one area the researchers are also currently exploring.

“We have shown that, with the right AI techniques, this simple stain can be very powerful. There is still much more research to do, but we need to take the organization of cells into account in more of our studies,” Uhler says.

This research was funded, in part, by the Eric and Wendy Schmidt Center at the Broad Institute, ETH Zurich, the Paul Scherrer Institute, the Swiss National Science Foundation, the U.S. National Institutes of Health, the U.S. Office of Naval Research, the MIT Jameel Clinic for Machine Learning and Health, the MIT-IBM Watson AI Lab, and a Simons Investigator Award.

When it comes to heating up the planet, not all greenhouse gases are created equal. They vary widely in their global warming potential (GWP), a measure of how much infrared thermal radiation a greenhouse gas would absorb over a given time frame once it enters the atmosphere. For example, measured over a 100-year period, the GWP of methane is about 28 times that of carbon dioxide (CO₂), and the GWPs of a class of greenhouse gases known as perfluorocarbons (PFCs) are thousands of times that of CO₂. The lifespans in the atmosphere of different greenhouse gases also vary widely. Methane persists in the atmosphere for around 10 years; CO₂ for over 100 years, and PFCs for up to tens of thousands of years.

Given the high GWPs and lifespans of PFCs, their emissions could pose a major roadblock to achieving the aspirational goal of the Paris Agreement on climate change — to limit the increase in global average surface temperature to 1.5 degrees Celsius above preindustrial levels. Now, two new studies based on atmospheric observations inside China and high-resolution atmospheric models show a rapid rise in Chinese emissions over the last decade (2011 to 2020 or 2021) of three PFCs: tetrafluoromethane (PFC-14) and hexafluoroethane (PFC-116) (results in PNAS), and perfluorocyclobutane (PFC-318) (results in Environmental Science & Technology).

Both studies find that Chinese emissions have played a dominant role in driving up global emission levels for all three PFCs.

The PNAS study identifies substantial PFC-14 and PFC-116 emission sources in the less-populated western regions of China from 2011 to 2021, likely due to the large amount of aluminum industry in these regions. The semiconductor industry also contributes to some of the emissions detected in the more economically developed eastern regions. These emissions are byproducts from aluminum smelting, or occur during the use of the two PFCs in the production of semiconductors and flat panel displays. During the observation period, emissions of both gases in China rose by 78 percent, accounting for most of the increase in global emissions of these gases.

The ES&T study finds that during 2011-20, a 70 percent increase in Chinese PFC-318 emissions (contributing more than half of the global emissions increase of this gas) — originated primarily in eastern China. The regions with high emissions of PFC-318 in China overlap with geographical areas densely populated with factories that produce polytetrafluoroethylene (PTFE, commonly used for nonstick cookware coatings), implying that PTFE factories are major sources of PFC-318 emissions in China. In these factories, PFC-318 is formed as a byproduct.

“Using atmospheric observations from multiple monitoring sites, we not only determined the magnitudes of PFC emissions, but also pinpointed the possible locations of their sources,” says Minde An, a postdoc at the MIT Center for Global Change Science (CGCS), and corresponding author of both studies. “Identifying the actual source industries contributing to these PFC emissions, and understanding the reasons for these largely byproduct emissions, can provide guidance for developing region- or industry-specific mitigation strategies.”

“These three PFCs are largely produced as unwanted byproducts during the manufacture of otherwise widely used industrial products,” says MIT professor of atmospheric sciences Ronald Prinn, director of both the MIT Joint Program on the Science and Policy of Global Change and CGCS, and a co-author of both studies. “Phasing out emissions of PFCs as early as possible is highly beneficial for achieving global climate mitigation targets and is likely achievable by recycling programs and targeted technological improvements in these industries.”

Findings in both studies were obtained, in part, from atmospheric observations collected from nine stations within a Chinese network, including one station from the Advanced Global Atmospheric Gases Experiment (AGAGE) network. For comparison, global total emissions were determined from five globally distributed, relatively unpolluted “background” AGAGE stations, as reported in the latest United Nations Environment Program and World Meteorological Organization Ozone Assessment report.

The concept of short-range order (SRO) — the arrangement of atoms over small distances — in metallic alloys has been underexplored in materials science and engineering. But the past decade has seen renewed interest in quantifying it, since decoding SRO is a crucial step toward developing tailored high-performing alloys, such as stronger or heat-resistant materials.

Understanding how atoms arrange themselves is no easy task and must be verified using intensive lab experiments or computer simulations based on imperfect models. These hurdles have made it difficult to fully explore SRO in metallic alloys.

But Killian Sheriff and Yifan Cao, graduate students in MIT’s Department of Materials Science and Engineering (DMSE), are using machine learning to quantify, atom-by-atom, the complex chemical arrangements that make up SRO. Under the supervision of Assistant Professor Rodrigo Freitas, and with the help of Assistant Professor Tess Smidt in the Department of Electrical Engineering and Computer Science, their work was recently published in The Proceedings of the National Academy of Sciences.

Interest in understanding SRO is linked to the excitement around advanced materials called high-entropy alloys, whose complex compositions give them superior properties.

Typically, materials scientists develop alloys by using one element as a base and adding small quantities of other elements to enhance specific properties. The addition of chromium to nickel, for example, makes the resulting metal more resistant to corrosion.

Unlike most traditional alloys, high-entropy alloys have several elements, from three up to 20, in nearly equal proportions. This offers a vast design space. “It’s like you’re making a recipe with a lot more ingredients,” says Cao.

The goal is to use SRO as a “knob” to tailor material properties by mixing chemical elements in high-entropy alloys in unique ways. This approach has potential applications in industries such as aerospace, biomedicine, and electronics, driving the need to explore permutations and combinations of elements, Cao says.

Capturing short-range order

Short-range order refers to the tendency of atoms to form chemical arrangements with specific neighboring atoms. While a superficial look at an alloy’s elemental distribution might indicate that its constituent elements are randomly arranged, it is often not so. “Atoms have a preference for having specific neighboring atoms arranged in particular patterns,” Freitas says. “How often these patterns arise and how they are distributed in space is what defines SRO.”

Understanding SRO unlocks the keys to the kingdom of high-entropy materials. Unfortunately, not much is known about SRO in high-entropy alloys. “It’s like we’re trying to build a huge Lego model without knowing what’s the smallest piece of Lego that you can have,” says Sheriff.

Traditional methods for understanding SRO involve small computational models, or simulations with a limited number of atoms, providing an incomplete picture of complex material systems. “High-entropy materials are chemically complex — you can’t simulate them well with just a few atoms; you really need to go a few length scales above that to capture the material accurately,” Sheriff says. “Otherwise, it’s like trying to understand your family tree without knowing one of the parents.”

SRO has also been calculated by using basic mathematics, counting immediate neighbors for a few atoms and computing what that distribution might look like on average. Despite its popularity, the approach has limitations, as it offers an incomplete picture of SRO.

Fortunately, researchers are leveraging machine learning to overcome the shortcomings of traditional approaches for capturing and quantifying SRO.

Hyunseok Oh, assistant professor in the Department of Materials Science and Engineering at the University of Wisconsin at Madison and a former DMSE postdoc, is excited about investigating SRO more fully. Oh, who was not involved in this study, explores how to leverage alloy composition, processing methods, and their relationship to SRO to design better alloys. “The physics of alloys and the atomistic origin of their properties depend on short-range ordering, but the accurate calculation of short-range ordering has been almost impossible,” says Oh. 

A two-pronged machine learning solution

To study SRO using machine learning, it helps to picture the crystal structure in high-entropy alloys as a connect-the-dots game in an coloring book, Cao says.

“You need to know the rules for connecting the dots to see the pattern.” And you need to capture the atomic interactions with a simulation that is big enough to fit the entire pattern. 

First, understanding the rules meant reproducing the chemical bonds in high-entropy alloys. “There are small energy differences in chemical patterns that lead to differences in short-range order, and we didn’t have a good model to do that,” Freitas says. The model the team developed is the first building block in accurately quantifying SRO.

The second part of the challenge, ensuring that researchers get the whole picture, was more complex. High-entropy alloys can exhibit billions of chemical “motifs,” combinations of arrangements of atoms. Identifying these motifs from simulation data is difficult because they can appear in symmetrically equivalent forms — rotated, mirrored, or inverted. At first glance, they may look different but still contain the same chemical bonds.

The team solved this problem by employing 3D Euclidean neural networks. These advanced computational models allowed the researchers to identify chemical motifs from simulations of high-entropy materials with unprecedented detail, examining them atom-by-atom.

The final task was to quantify the SRO. Freitas used machine learning to evaluate the different chemical motifs and tag each with a number. When researchers want to quantify the SRO for a new material, they run it by the model, which sorts it in its database and spits out an answer.

The team also invested additional effort in making their motif identification framework more accessible. “We have this sheet of all possible permutations of [SRO] already set up, and we know what number each of them got through this machine learning process,” Freitas says. “So later, as we run into simulations, we can sort them out to tell us what that new SRO will look like.” The neural network easily recognizes symmetry operations and tags equivalent structures with the same number.

“If you had to compile all the symmetries yourself, it’s a lot of work. Machine learning organized this for us really quickly and in a way that was cheap enough that we could apply it in practice,” Freitas says.

Enter the world’s fastest supercomputer

This summer, Cao and Sheriff and team will have a chance to explore how SRO can change under routine metal processing conditions, like casting and cold-rolling, through the U.S. Department of Energy’s INCITE program, which allows access to Frontier, the world’s fastest supercomputer.

“If you want to know how short-range order changes during the actual manufacturing of metals, you need to have a very good model and a very large simulation,” Freitas says. The team already has a strong model; it will now leverage INCITE’s computing facilities for the robust simulations required.

“With that we expect to uncover the sort of mechanisms that metallurgists could employ to engineer alloys with pre-determined SRO,” Freitas adds.

Sheriff is excited about the research’s many promises. One is the 3D information that can be obtained about chemical SRO. Whereas traditional transmission electron microscopes and other methods are limited to two-dimensional data, physical simulations can fill in the dots and give full access to 3D information, Sheriff says.

“We have introduced a framework to start talking about chemical complexity,” Sheriff explains. “Now that we can understand this, there’s a whole body of materials science on classical alloys to develop predictive tools for high-entropy materials.”

That could lead to the purposeful design of new classes of materials instead of simply shooting in the dark.

The research was funded by the MathWorks Ignition Fund, MathWorks Engineering Fellowship Fund, and the Portuguese Foundation for International Cooperation in Science, Technology and Higher Education in the MIT–Portugal Program.

Hot Jupiters are some of the most extreme planets in the galaxy. These scorching worlds are as massive as Jupiter, and they swing wildly close to their star, whirling around in a few days compared to our own gas giant’s leisurely 4,000-day orbit around the sun.

Scientists suspect, though, that hot Jupiters weren’t always so hot and in fact may have formed as “cold Jupiters,” in more frigid, distant environs. But how they evolved to be the star-hugging gas giants that astronomers observe today is a big unknown.

Now, astronomers at MIT, Penn State University, and elsewhere have discovered a hot Jupiter “progenitor” — a sort of juvenile planet that is in the midst of becoming a hot Jupiter. And its orbit is providing some answers to how hot Jupiters evolve.

The new planet, which astronomers labeled TIC 241249530 b, orbits a star that is about 1,100 light-years from Earth. The planet circles its star in a highly “eccentric” orbit, meaning that it comes extremely close to the star before slinging far out, then doubling back, in a narrow, elliptical circuit. If the planet was part of our solar system, it would come 10 times closer to the sun than Mercury, before hurtling out, just past Earth, then back around. By the scientists’ estimates, the planet’s stretched-out orbit has the highest eccentricity of any planet detected to date.

The new planet’s orbit is also unique in its “retrograde” orientation. Unlike the Earth and other planets in the solar system, which orbit in the same direction as the sun spins, the new planet travels in a direction that is counter to its star’s rotation.

The team ran simulations of orbital dynamics and found that the planet’s highly eccentric and retrograde orbit are signs that it is likely evolving into a hot Jupiter, through “high-eccentricity migration” — a process by which a planet’s orbit wobbles and progressively shrinks as it interacts with another star or planet on a much wider orbit.

In the case of TIC 241249530 b, the researchers determined that the planet orbits around a primary star that itself orbits around a secondary star, as part of a stellar binary system. The interactions between the two orbits — of the planet and its star — have caused the planet to gradually migrate closer to its star over time.

The planet’s orbit is currently elliptical in shape, and the planet takes about 167 days to complete a lap around its star. The researchers predict that in 1 billion years, the planet will migrate into a much tighter, circular orbit, when it will then circle its star every few days. At that point, the planet will have fully evolved into a hot Jupiter.

“This new planet supports the theory that high eccentricity migration should account for some fraction of hot Jupiters,” says Sarah Millholland, assistant professor of physics in MIT’s Kavli Institute for Astrophysics and Space Research. “We think that when this planet formed, it would have been a frigid world. And because of the dramatic orbital dynamics, it will become a hot Jupiter in about a billion years, with temperatures of several thousand kelvin. So it’s a huge shift from where it started.”

Millholland and her colleagues have published their findings today in the journal Nature. Her co-authors are MIT undergraduate Haedam Im, lead author Arvind Gupta of Penn State University and NSF NOIRLab, and collaborators at multiple other universities, institutions, and observatories.

“Radical seasons”

The new planet was first spotted in data taken by NASA’s Transiting Exoplanet Survey Satellite (TESS), an MIT-led mission that monitors the brightness of nearby stars for “transits,” or brief dips in starlight that could signal the presence of a planet passing in front of, and temporarily blocking, a star’s light.

On Jan. 12, 2020, TESS picked up a possible transit of the star TIC 241249530. Gupta and his colleagues at Penn State determined that the transit was consistent with a Jupiter-sized planet crossing in front of the star. They then acquired measurements from other observatories of the star’s radial velocity, which estimates a star’s wobble, or the degree to which it moves back and forth, in response to other nearby objects that might gravitationally tug on the star.

Those measurements confirmed that a Jupiter-sized planet was orbiting the star and that its orbit was highly eccentric, bringing the planet extremely close to the star before flinging it far out.

Prior to this detection, astronomers had known of only one other planet, HD 80606 b, that was thought to be an early hot Jupiter. That planet, discovered in 2001, held the record for having the highest eccentricity, until now.

“This new planet experiences really dramatic changes in starlight throughout its orbit,” Millholland says. “There must be really radical seasons and an absolutely scorched atmosphere every time it passes close to the star.”

“Dance of orbits”

How could a planet have fallen into such an extreme orbit? And how might its eccentricity evolve over time? For answers, Im and Millholland ran simulations of planetary orbital dynamics to model how the planet may have evolved throughout its history and how it might carry on over hundreds of millions of years.

The team modeled the gravitational interactions between the planet, its star, and the second nearby star. Gupta and his colleagues had observed that the two stars orbit each other in a binary system, while the planet is simultaneously orbiting the closer star. The configuration of the two orbits is somewhat like a circus performer twirling a hula hoop around her waist, while spinning a second hula hoop around her wrist.

Millholland and Im ran multiple simulations, each with a different set of starting conditions, to see which condition, when run forward over several billions of years, produced the configuration of planetary and stellar orbits that Gupta’s team observed in the present day. They then ran the best match even further into the future to predict how the system will evolve over the next several billion years.

These simulations revealed that the new planet is likely in the midst of evolving into a hot Jupiter: Several billion years ago, the planet formed as a cold Jupiter, far from its star, in a region cold enough to condense and take shape. Newly formed, the planet likely orbited the star in a circular path. This conventional orbit, however, gradually stretched and grew eccentric, as it experienced gravitational forces from the star’s misaligned orbit with its second, binary star.

“It’s a pretty extreme process in that the changes to the planet’s orbit are massive,” Millholland says. “It’s a big dance of orbits that’s happening over billions of years, and the planet’s just going along for the ride.”

In another billion years, the simulations show that the planet’s orbit will stabilize in a close-in, circular path around its star.

“Then, the planet will fully become a hot Jupiter,” Millholland says.

The team’s observations, along with their simulations of the planet’s evolution, support the theory that hot Jupiters can form through high eccentricity migration, a process by which a planet gradually moves into place via extreme changes to its orbit over time.

“It’s clear not only from this, but other statistical studies too, that high eccentricity migration should account for some fraction of hot Jupiters,” Millholland notes. “This system highlights how incredibly diverse exoplanets can be. They are mysterious other worlds that can have wild orbits that tell a story of how they got that way and where they’re going. For this planet, it’s not quite finished its journey yet.”

“It is really hard to catch these hot Jupiter progenitors ‘in the act’ as they undergo their super eccentric episodes, so it is very exciting to find a system that undergoes this process,” says Smadar Naoz, a professor of physics and astronomy at the University of California at Los Angeles, who was not involved with the study. “I believe that this discovery opens the door to a deeper understanding of the birth configuration of the exoplanetary system.”

Neural networks have made a seismic impact on how engineers design controllers for robots, catalyzing more adaptive and efficient machines. Still, these brain-like machine-learning systems are a double-edged sword: Their complexity makes them powerful, but it also makes it difficult to guarantee that a robot powered by a neural network will safely accomplish its task.

The traditional way to verify safety and stability is through techniques called Lyapunov functions. If you can find a Lyapunov function whose value consistently decreases, then you can know that unsafe or unstable situations associated with higher values will never happen. For robots controlled by neural networks, though, prior approaches for verifying Lyapunov conditions didn’t scale well to complex machines.

Researchers from MIT’s Computer Science and Artificial Intelligence Laboratory (CSAIL) and elsewhere have now developed new techniques that rigorously certify Lyapunov calculations in more elaborate systems. Their algorithm efficiently searches for and verifies a Lyapunov function, providing a stability guarantee for the system. This approach could potentially enable safer deployment of robots and autonomous vehicles, including aircraft and spacecraft.

To outperform previous algorithms, the researchers found a frugal shortcut to the training and verification process. They generated cheaper counterexamples — for example, adversarial data from sensors that could’ve thrown off the controller — and then optimized the robotic system to account for them. Understanding these edge cases helped machines learn how to handle challenging circumstances, which enabled them to operate safely in a wider range of conditions than previously possible. Then, they developed a novel verification formulation that enables the use of a scalable neural network verifier, α,β-CROWN, to provide rigorous worst-case scenario guarantees beyond the counterexamples.

“We’ve seen some impressive empirical performances in AI-controlled machines like humanoids and robotic dogs, but these AI controllers lack the formal guarantees that are crucial for safety-critical systems,” says Lujie Yang, MIT electrical engineering and computer science (EECS) PhD student and CSAIL affiliate who is a co-lead author of a new paper on the project alongside Toyota Research Institute researcher Hongkai Dai SM ’12, PhD ’16. “Our work bridges the gap between that level of performance from neural network controllers and the safety guarantees needed to deploy more complex neural network controllers in the real world,” notes Yang.

For a digital demonstration, the team simulated how a quadrotor drone with lidar sensors would stabilize in a two-dimensional environment. Their algorithm successfully guided the drone to a stable hover position, using only the limited environmental information provided by the lidar sensors. In two other experiments, their approach enabled the stable operation of two simulated robotic systems over a wider range of conditions: an inverted pendulum and a path-tracking vehicle. These experiments, though modest, are relatively more complex than what the neural network verification community could have done before, especially because they included sensor models.

“Unlike common machine learning problems, the rigorous use of neural networks as Lyapunov functions requires solving hard global optimization problems, and thus scalability is the key bottleneck,” says Sicun Gao, associate professor of computer science and engineering at the University of California at San Diego, who wasn’t involved in this work. “The current work makes an important contribution by developing algorithmic approaches that are much better tailored to the particular use of neural networks as Lyapunov functions in control problems. It achieves impressive improvement in scalability and the quality of solutions over existing approaches. The work opens up exciting directions for further development of optimization algorithms for neural Lyapunov methods and the rigorous use of deep learning in control and robotics in general.”

Yang and her colleagues’ stability approach has potential wide-ranging applications where guaranteeing safety is crucial. It could help ensure a smoother ride for autonomous vehicles, like aircraft and spacecraft. Likewise, a drone delivering items or mapping out different terrains could benefit from such safety guarantees.

The techniques developed here are very general and aren’t just specific to robotics; the same techniques could potentially assist with other applications, such as biomedicine and industrial processing, in the future.

While the technique is an upgrade from prior works in terms of scalability, the researchers are exploring how it can perform better in systems with higher dimensions. They’d also like to account for data beyond lidar readings, like images and point clouds.

As a future research direction, the team would like to provide the same stability guarantees for systems that are in uncertain environments and subject to disturbances. For instance, if a drone faces a strong gust of wind, Yang and her colleagues want to ensure it’ll still fly steadily and complete the desired task. 

Also, they intend to apply their method to optimization problems, where the goal would be to minimize the time and distance a robot needs to complete a task while remaining steady. They plan to extend their technique to humanoids and other real-world machines, where a robot needs to stay stable while making contact with its surroundings.

Russ Tedrake, the Toyota Professor of EECS, Aeronautics and Astronautics, and Mechanical Engineering at MIT, vice president of robotics research at TRI, and CSAIL member, is a senior author of this research. The paper also credits University of California at Los Angeles PhD student Zhouxing Shi and associate professor Cho-Jui Hsieh, as well as University of Illinois Urbana-Champaign assistant professor Huan Zhang. Their work was supported, in part, by Amazon, the National Science Foundation, the Office of Naval Research, and the AI2050 program at Schmidt Sciences. The researchers’ paper will be presented at the 2024 International Conference on Machine Learning.

It is estimated that about 70 percent of the energy generated worldwide ends up as waste heat.

If scientists could better predict how heat moves through semiconductors and insulators, they could design more efficient power generation systems. However, the thermal properties of materials can be exceedingly difficult to model.

The trouble comes from phonons, which are subatomic particles that carry heat. Some of a material’s thermal properties depend on a measurement called the phonon dispersion relation, which can be incredibly hard to obtain, let alone utilize in the design of a system.

A team of researchers from MIT and elsewhere tackled this challenge by rethinking the problem from the ground up. The result of their work is a new machine-learning framework that can predict phonon dispersion relations up to 1,000 times faster than other AI-based techniques, with comparable or even better accuracy. Compared to more traditional, non-AI-based approaches, it could be 1 million times faster.

This method could help engineers design energy generation systems that produce more power, more efficiently. It could also be used to develop more efficient microelectronics, since managing heat remains a major bottleneck to speeding up electronics.

“Phonons are the culprit for the thermal loss, yet obtaining their properties is notoriously challenging, either computationally or experimentally,” says Mingda Li, associate professor of nuclear science and engineering and senior author of a paper on this technique.

Li is joined on the paper by co-lead authors Ryotaro Okabe, a chemistry graduate student; and Abhijatmedhi Chotrattanapituk, an electrical engineering and computer science graduate student; Tommi Jaakkola, the Thomas Siebel Professor of Electrical Engineering and Computer Science at MIT; as well as others at MIT, Argonne National Laboratory, Harvard University, the University of South Carolina, Emory University, the University of California at Santa Barbara, and Oak Ridge National Laboratory. The research appears in Nature Computational Science.

Predicting phonons

Heat-carrying phonons are tricky to predict because they have an extremely wide frequency range, and the particles interact and travel at different speeds.

A material’s phonon dispersion relation is the relationship between energy and momentum of phonons in its crystal structure. For years, researchers have tried to predict phonon dispersion relations using machine learning, but there are so many high-precision calculations involved that models get bogged down.

“If you have 100 CPUs and a few weeks, you could probably calculate the phonon dispersion relation for one material. The whole community really wants a more efficient way to do this,” says Okabe.

The machine-learning models scientists often use for these calculations are known as graph neural networks (GNN). A GNN converts a material’s atomic structure into a crystal graph comprising multiple nodes, which represent atoms, connected by edges, which represent the interatomic bonding between atoms.

While GNNs work well for calculating many quantities, like magnetization or electrical polarization, they are not flexible enough to efficiently predict an extremely high-dimensional quantity like the phonon dispersion relation. Because phonons can travel around atoms on X, Y, and Z axes, their momentum space is hard to model with a fixed graph structure.

To gain the flexibility they needed, Li and his collaborators devised virtual nodes.

They create what they call a virtual node graph neural network (VGNN) by adding a series of flexible virtual nodes to the fixed crystal structure to represent phonons. The virtual nodes enable the output of the neural network to vary in size, so it is not restricted by the fixed crystal structure.

Virtual nodes are connected to the graph in such a way that they can only receive messages from real nodes. While virtual nodes will be updated as the model updates real nodes during computation, they do not affect the accuracy of the model.

“The way we do this is very efficient in coding. You just generate a few more nodes in your GNN. The physical location doesn’t matter, and the real nodes don’t even know the virtual nodes are there,” says Chotrattanapituk.

Cutting out complexity

Since it has virtual nodes to represent phonons, the VGNN can skip many complex calculations when estimating phonon dispersion relations, which makes the method more efficient than a standard GNN. 

The researchers proposed three different versions of VGNNs with increasing complexity. Each can be used to predict phonons directly from a material’s atomic coordinates.

Because their approach has the flexibility to rapidly model high-dimensional properties, they can use it to estimate phonon dispersion relations in alloy systems. These complex combinations of metals and nonmetals are especially challenging for traditional approaches to model.

The researchers also found that VGNNs offered slightly greater accuracy when predicting a material’s heat capacity. In some instances, prediction errors were two orders of magnitude lower with their technique.

A VGNN could be used to calculate phonon dispersion relations for a few thousand materials in just a few seconds with a personal computer, Li says.

This efficiency could enable scientists to search a larger space when seeking materials with certain thermal properties, such as superior thermal storage, energy conversion, or superconductivity.

Moreover, the virtual node technique is not exclusive to phonons, and could also be used to predict challenging optical and magnetic properties.

In the future, the researchers want to refine the technique so virtual nodes have greater sensitivity to capture small changes that can affect phonon structure.

“Researchers got too comfortable using graph nodes to represent atoms, but we can rethink that. Graph nodes can be anything. And virtual nodes are a very generic approach you could use to predict a lot of high-dimensional quantities,” Li says.

“The authors’ innovative approach significantly augments the graph neural network description of solids by incorporating key physics-informed elements through virtual nodes, for instance, informing wave-vector dependent band-structures and dynamical matrices,” says Olivier Delaire, associate professor in the Thomas Lord Department of Mechanical Engineering and Materials Science at Duke University, who was not involved with this work. “I find that the level of acceleration in predicting complex phonon properties is amazing, several orders of magnitude faster than a state-of-the-art universal machine-learning interatomic potential. Impressively, the advanced neural net captures fine features and obeys physical rules. There is great potential to expand the model to describe other important material properties: Electronic, optical, and magnetic spectra and band structures come to mind.”

This work is supported by the U.S. Department of Energy, National Science Foundation, a Mathworks Fellowship, a Sow-Hsin Chen Fellowship, the Harvard Quantum Initiative, and the Oak Ridge National Laboratory.

Foundation models are massive deep-learning models that have been pretrained on an enormous amount of general-purpose, unlabeled data. They can be applied to a variety of tasks, like generating images or answering customer questions.

But these models, which serve as the backbone for powerful artificial intelligence tools like ChatGPT and DALL-E, can offer up incorrect or misleading information. In a safety-critical situation, such as a pedestrian approaching a self-driving car, these mistakes could have serious consequences.

To help prevent such mistakes, researchers from MIT and the MIT-IBM Watson AI Lab developed a technique to estimate the reliability of foundation models before they are deployed to a specific task.

They do this by considering a set of foundation models that are slightly different from one another. Then they use their algorithm to assess the consistency of the representations each model learns about the same test data point. If the representations are consistent, it means the model is reliable.

When they compared their technique to state-of-the-art baseline methods, it was better at capturing the reliability of foundation models on a variety of downstream classification tasks.

Someone could use this technique to decide if a model should be applied in a certain setting, without the need to test it on a real-world dataset. This could be especially useful when datasets may not be accessible due to privacy concerns, like in health care settings. In addition, the technique could be used to rank models based on reliability scores, enabling a user to select the best one for their task.

“All models can be wrong, but models that know when they are wrong are more useful. The problem of quantifying uncertainty or reliability is more challenging for these foundation models because their abstract representations are difficult to compare. Our method allows one to quantify how reliable a representation model is for any given input data,” says senior author Navid Azizan, the Esther and Harold E. Edgerton Assistant Professor in the MIT Department of Mechanical Engineering and the Institute for Data, Systems, and Society (IDSS), and a member of the Laboratory for Information and Decision Systems (LIDS).

He is joined on a paper about the work by lead author Young-Jin Park, a LIDS graduate student; Hao Wang, a research scientist at the MIT-IBM Watson AI Lab; and Shervin Ardeshir, a senior research scientist at Netflix. The paper will be presented at the Conference on Uncertainty in Artificial Intelligence.

Measuring consensus

Traditional machine-learning models are trained to perform a specific task. These models typically make a concrete prediction based on an input. For instance, the model might tell you whether a certain image contains a cat or a dog. In this case, assessing reliability could be a matter of looking at the final prediction to see if the model is right.

But foundation models are different. The model is pretrained using general data, in a setting where its creators don’t know all downstream tasks it will be applied to. Users adapt it to their specific tasks after it has already been trained.

Unlike traditional machine-learning models, foundation models don’t give concrete outputs like “cat” or “dog” labels. Instead, they generate an abstract representation based on an input data point.

To assess the reliability of a foundation model, the researchers used an ensemble approach by training several models which share many properties but are slightly different from one another.

“Our idea is like measuring the consensus. If all those foundation models are giving consistent representations for any data in our dataset, then we can say this model is reliable,” Park says.

But they ran into a problem: How could they compare abstract representations?

“These models just output a vector, comprised of some numbers, so we can’t compare them easily,” he adds.

They solved this problem using an idea called neighborhood consistency.

For their approach, the researchers prepare a set of reliable reference points to test on the ensemble of models. Then, for each model, they investigate the reference points located near that model’s representation of the test point.

By looking at the consistency of neighboring points, they can estimate the reliability of the models.

Aligning the representations

Foundation models map data points to what is known as a representation space. One way to think about this space is as a sphere. Each model maps similar data points to the same part of its sphere, so images of cats go in one place and images of dogs go in another.

But each model would map animals differently in its own sphere, so while cats may be grouped near the South Pole of one sphere, another model could map cats somewhere in the Northern Hemisphere.

The researchers use the neighboring points like anchors to align those spheres so they can make the representations comparable. If a data point’s neighbors are consistent across multiple representations, then one should be confident about the reliability of the model’s output for that point.

When they tested this approach on a wide range of classification tasks, they found that it was much more consistent than baselines. Plus, it wasn’t tripped up by challenging test points that caused other methods to fail.

Moreover, their approach can be used to assess reliability for any input data, so one could evaluate how well a model works for a particular type of individual, such as a patient with certain characteristics.

“Even if the models all have average performance overall, from an individual point of view, you’d prefer the one that works best for that individual,” Wang says.

However, one limitation comes from the fact that they must train an ensemble of foundation models, which is computationally expensive. In the future, they plan to find more efficient ways to build multiple models, perhaps by using small perturbations of a single model.

“With the current trend of using foundational models for their embeddings to support various downstream tasks — from fine-tuning to retrieval augmented generation — the topic of quantifying uncertainty at the representation level is increasingly important, but challenging, as embeddings on their own have no grounding. What matters instead is how embeddings of different inputs are related to one another, an idea that this work neatly captures through the proposed neighborhood consistency score,” says Marco Pavone, an associate professor in the Department of Aeronautics and Astronautics at Stanford University, who was not involved with this work. “This is a promising step towards high quality uncertainty quantifications for embedding models, and I’m excited to see future extensions which can operate without requiring model-ensembling to really enable this approach to scale to foundation-size models.”

This work is funded, in part, by the MIT-IBM Watson AI Lab, MathWorks, and Amazon.

There are many drugs that anesthesiologists can use to induce unconsciousness in patients. Exactly how these drugs cause the brain to lose consciousness has been a longstanding question, but MIT neuroscientists have now answered that question for one commonly used anesthesia drug.

Using a novel technique for analyzing neuron activity, the researchers discovered that the drug propofol induces unconsciousness by disrupting the brain’s normal balance between stability and excitability. The drug causes brain activity to become increasingly unstable, until the brain loses consciousness.

“The brain has to operate on this knife’s edge between excitability and chaos. It’s got to be excitable enough for its neurons to influence one another, but if it gets too excitable, it spins off into chaos. Propofol seems to disrupt the mechanisms that keep the brain in that narrow operating range,” says Earl K. Miller, the Picower Professor of Neuroscience and a member of MIT’s Picower Institute for Learning and Memory.

The new findings, reported today in Neuron, could help researchers develop better tools for monitoring patients as they undergo general anesthesia.

Miller and Ila Fiete, a professor of brain and cognitive sciences, the director of the K. Lisa Yang Integrative Computational Neuroscience Center (ICoN), and a member of MIT’s McGovern Institute for Brain Research, are the senior authors of the new study. MIT graduate student Adam Eisen and MIT postdoc Leo Kozachkov are the lead authors of the paper.

Losing consciousness

Propofol is a drug that binds to GABA receptors in the brain, inhibiting neurons that have those receptors. Other anesthesia drugs act on different types of receptors, and the mechanism for how all of these drugs produce unconsciousness is not fully understood.

Miller, Fiete, and their students hypothesized that propofol, and possibly other anesthesia drugs, interfere with a brain state known as “dynamic stability.” In this state, neurons have enough excitability to respond to new input, but the brain is able to quickly regain control and prevent them from becoming overly excited.

Previous studies of how anesthesia drugs affect this balance have found conflicting results: Some suggested that during anesthesia, the brain shifts toward becoming too stable and unresponsive, which leads to loss of consciousness. Others found that the brain becomes too excitable, leading to a chaotic state that results in unconsciousness.

Part of the reason for these conflicting results is that it has been difficult to accurately measure dynamic stability in the brain. Measuring dynamic stability as consciousness is lost would help researchers determine if unconsciousness results from too much stability or too little stability.

In this study, the researchers analyzed electrical recordings made in the brains of animals that received propofol over an hour-long period, during which they gradually lost consciousness. The recordings were made in four areas of the brain that are involved in vision, sound processing, spatial awareness, and executive function.

These recordings covered only a tiny fraction of the brain’s overall activity, so to overcome that, the researchers used a technique called delay embedding. This technique allows researchers to characterize dynamical systems from limited measurements by augmenting each measurement with measurements that were recorded previously.

Using this method, the researchers were able to quantify how the brain responds to sensory inputs, such as sounds, or to spontaneous perturbations of neural activity.

In the normal, awake state, neural activity spikes after any input, then returns to its baseline activity level. However, once propofol dosing began, the brain started taking longer to return to its baseline after these inputs, remaining in an overly excited state. This effect became more and more pronounced until the animals lost consciousness.

This suggests that propofol’s inhibition of neuron activity leads to escalating instability, which causes the brain to lose consciousness, the researchers say.

Better anesthesia control

To see if they could replicate this effect in a computational model, the researchers created a simple neural network. When they increased the inhibition of certain nodes in the network, as propofol does in the brain, network activity became destabilized, similar to the unstable activity the researchers saw in the brains of animals that received propofol.

“We looked at a simple circuit model of interconnected neurons, and when we turned up inhibition in that, we saw a destabilization. So, one of the things we’re suggesting is that an increase in inhibition can generate instability, and that is subsequently tied to loss of consciousness,” Eisen says.

As Fiete explains, “This paradoxical effect, in which boosting inhibition destabilizes the network rather than silencing or stabilizing it, occurs because of disinhibition. When propofol boosts the inhibitory drive, this drive inhibits other inhibitory neurons, and the result is an overall increase in brain activity.”

The researchers suspect that other anesthetic drugs, which act on different types of neurons and receptors, may converge on the same effect through different mechanisms — a possibility that they are now exploring.

If this turns out to be true, it could be helpful to the researchers’ ongoing efforts to develop ways to more precisely control the level of anesthesia that a patient is experiencing. These systems, which Miller is working on with Emery Brown, the Edward Hood Taplin Professor of Medical Engineering at MIT, work by measuring the brain’s dynamics and then adjusting drug dosages accordingly, in real-time.

“If you find common mechanisms at work across different anesthetics, you can make them all safer by tweaking a few knobs, instead of having to develop safety protocols for all the different anesthetics one at a time,” Miller says. “You don’t want a different system for every anesthetic they’re going to use in the operating room. You want one that’ll do it all.”

The researchers also plan to apply their technique for measuring dynamic stability to other brain states, including neuropsychiatric disorders.

“This method is pretty powerful, and I think it’s going to be very exciting to apply it to different brain states, different types of anesthetics, and also other neuropsychiatric conditions like depression and schizophrenia,” Fiete says.

The research was funded by the Office of Naval Research, the National Institute of Mental Health, the National Institute of Neurological Disorders and Stroke, the National Science Foundation Directorate for Computer and Information Science and Engineering, the Simons Center for the Social Brain, the Simons Collaboration on the Global Brain, the JPB Foundation, the McGovern Institute, and the Picower Institute. 

When it comes to artificial intelligence, appearances can be deceiving. The mystery surrounding the inner workings of large language models (LLMs) stems from their vast size, complex training methods, hard-to-predict behaviors, and elusive interpretability.

MIT's Computer Science and Artificial Intelligence Laboratory (CSAIL) researchers recently peered into the proverbial magnifying glass to examine how LLMs fare with variations of different tasks, revealing intriguing insights into the interplay between memorization and reasoning skills. It turns out that their reasoning abilities are often overestimated.

The study compared “default tasks,” the common tasks a model is trained and tested on, with “counterfactual scenarios,” hypothetical situations deviating from default conditions — which models like GPT-4 and Claude can usually be expected to cope with. The researchers developed some tests outside the models’ comfort zones by tweaking existing tasks instead of creating entirely new ones. They used a variety of datasets and benchmarks specifically tailored to different aspects of the models' capabilities for things like arithmetic, chess, evaluating code, answering logical questions, etc.

When users interact with language models, any arithmetic is usually in base-10, the familiar number base to the models. But observing that they do well on base-10 could give us a false impression of them having strong competency in addition. Logically, if they truly possess good addition skills, you’d expect reliably high performance across all number bases, similar to calculators or computers. Indeed, the research showed that these models are not as robust as many initially think. Their high performance is limited to common task variants and suffer from consistent and severe performance drop in the unfamiliar counterfactual scenarios, indicating a lack of generalizable addition ability. 

The pattern held true for many other tasks like musical chord fingering, spatial reasoning, and even chess problems where the starting positions of pieces were slightly altered. While human players are expected to still be able to determine the legality of moves in altered scenarios (given enough time), the models struggled and couldn’t perform better than random guessing, meaning they have limited ability to generalize to unfamiliar situations. And much of their performance on the standard tasks is likely not due to general task abilities, but overfitting to, or directly memorizing from, what they have seen in their training data.

“We’ve uncovered a fascinating aspect of large language models: they excel in familiar scenarios, almost like a well-worn path, but struggle when the terrain gets unfamiliar. This insight is crucial as we strive to enhance these models’ adaptability and broaden their application horizons,” says Zhaofeng Wu, an MIT PhD student in electrical engineering and computer science, CSAIL affiliate, and the lead author on a new paper about the research. “As AI is becoming increasingly ubiquitous in our society, it must reliably handle diverse scenarios, whether familiar or not. We hope these insights will one day inform the design of future LLMs with improved robustness.”

Despite the insights gained, there are, of course, limitations. The study’s focus on specific tasks and settings didn’t capture the full range of challenges the models could potentially encounter in real-world applications, signaling the need for more diverse testing environments. Future work could involve expanding the range of tasks and counterfactual conditions to uncover more potential weaknesses. This could mean looking at more complex and less common scenarios. The team also wants to improve interpretability by creating methods to better comprehend the rationale behind the models’ decision-making processes.

“As language models scale up, understanding their training data becomes increasingly challenging even for open models, let alone proprietary ones,” says Hao Peng, assistant professor at the University of Illinois at Urbana-Champaign. “The community remains puzzled about whether these models genuinely generalize to unseen tasks, or seemingly succeed by memorizing the training data. This paper makes important strides in addressing this question. It constructs a suite of carefully designed counterfactual evaluations, providing fresh insights into the capabilities of state-of-the-art LLMs. It reveals that their ability to solve unseen tasks is perhaps far more limited than anticipated by many. It has the potential to inspire future research towards identifying the failure modes of today’s models and developing better ones.”

Additional authors include Najoung Kim, who is a Boston University assistant professor and Google visiting researcher, and seven CSAIL affiliates: MIT electrical engineering and computer science (EECS) PhD students Linlu Qiu, Alexis Ross, Ekin Akyürek SM ’21, and Boyuan Chen; former postdoc and Apple AI/ML researcher Bailin Wang; and EECS assistant professors Jacob Andreas and Yoon Kim. 

The team’s study was supported, in part, by the MIT–IBM Watson AI Lab, the MIT Quest for Intelligence, and the National Science Foundation. The team presented the work at the North American Chapter of the Association for Computational Linguistics (NAACL) last month.

Because machine-learning models can give false predictions, researchers often equip them with the ability to tell a user how confident they are about a certain decision. This is especially important in high-stake settings, such as when models are used to help identify disease in medical images or filter job applications.

But a model’s uncertainty quantifications are only useful if they are accurate. If a model says it is 49 percent confident that a medical image shows a pleural effusion, then 49 percent of the time, the model should be right.

MIT researchers have introduced a new approach that can improve uncertainty estimates in machine-learning models. Their method not only generates more accurate uncertainty estimates than other techniques, but does so more efficiently.

In addition, because the technique is scalable, it can be applied to huge deep-learning models that are increasingly being deployed in health care and other safety-critical situations.

This technique could give end users, many of whom lack machine-learning expertise, better information they can use to determine whether to trust a model’s predictions or if the model should be deployed for a particular task.

“It is easy to see these models perform really well in scenarios where they are very good, and then assume they will be just as good in other scenarios. This makes it especially important to push this kind of work that seeks to better calibrate the uncertainty of these models to make sure they align with human notions of uncertainty,” says lead author Nathan Ng, a graduate student at the University of Toronto who is a visiting student at MIT.

Ng wrote the paper with Roger Grosse, an assistant professor of computer science at the University of Toronto; and senior author Marzyeh Ghassemi, an associate professor in the Department of Electrical Engineering and Computer Science and a member of the Institute of Medical Engineering Sciences and the Laboratory for Information and Decision Systems. The research will be presented at the International Conference on Machine Learning.

Quantifying uncertainty

Uncertainty quantification methods often require complex statistical calculations that don’t scale well to machine-learning models with millions of parameters. These methods also require users to make assumptions about the model and data used to train it.

The MIT researchers took a different approach. They use what is known as the minimum description length principle (MDL), which does not require the assumptions that can hamper the accuracy of other methods. MDL is used to better quantify and calibrate uncertainty for test points the model has been asked to label.

The technique the researchers developed, known as IF-COMP, makes MDL fast enough to use with the kinds of large deep-learning models deployed in many real-world settings.

MDL involves considering all possible labels a model could give a test point. If there are many alternative labels for this point that fit well, its confidence in the label it chose should decrease accordingly.

“One way to understand how confident a model is would be to tell it some counterfactual information and see how likely it is to believe you,” Ng says.

For example, consider a model that says a medical image shows a pleural effusion. If the researchers tell the model this image shows an edema, and it is willing to update its belief, then the model should be less confident in its original decision.

With MDL, if a model is confident when it labels a datapoint, it should use a very short code to describe that point. If it is uncertain about its decision because the point could have many other labels, it uses a longer code to capture these possibilities.

The amount of code used to label a datapoint is known as stochastic data complexity. If the researchers ask the model how willing it is to update its belief about a datapoint given contrary evidence, the stochastic data complexity should decrease if the model is confident.

But testing each datapoint using MDL would require an enormous amount of computation.

Speeding up the process

With IF-COMP, the researchers developed an approximation technique that can accurately estimate stochastic data complexity using a special function, known as an influence function. They also employed a statistical technique called temperature-scaling, which improves the calibration of the model’s outputs. This combination of influence functions and temperature-scaling enables high-quality approximations of the stochastic data complexity.

In the end, IF-COMP can efficiently produce well-calibrated uncertainty quantifications that reflect a model’s true confidence. The technique can also determine whether the model has mislabeled certain data points or reveal which data points are outliers.

The researchers tested their system on these three tasks and found that it was faster and more accurate than other methods.

“It is really important to have some certainty that a model is well-calibrated, and there is a growing need to detect when a specific prediction doesn’t look quite right. Auditing tools are becoming more necessary in machine-learning problems as we use large amounts of unexamined data to make models that will be applied to human-facing problems,” Ghassemi says.

IF-COMP is model-agnostic, so it can provide accurate uncertainty quantifications for many types of machine-learning models. This could enable it to be deployed in a wider range of real-world settings, ultimately helping more practitioners make better decisions.

“People need to understand that these systems are very fallible and can make things up as they go. A model may look like it is highly confident, but there are a ton of different things it is willing to believe given evidence to the contrary,” Ng says.

In the future, the researchers are interested in applying their approach to large language models and studying other potential use cases for the minimum description length principle. 

As container ships the size of city blocks cross the oceans to deliver cargo, their huge diesel engines emit large quantities of air pollutants that drive climate change and have human health impacts. It has been estimated that maritime shipping accounts for almost 3 percent of global carbon dioxide emissions and the industry’s negative impacts on air quality cause about 100,000 premature deaths each year.

Decarbonizing shipping to reduce these detrimental effects is a goal of the International Maritime Organization, a U.N. agency that regulates maritime transport. One potential solution is switching the global fleet from fossil fuels to sustainable fuels such as ammonia, which could be nearly carbon-free when considering its production and use.

But in a new study, an interdisciplinary team of researchers from MIT and elsewhere caution that burning ammonia for maritime fuel could worsen air quality further and lead to devastating public health impacts, unless it is adopted alongside strengthened emissions regulations.

Ammonia combustion generates nitrous oxide (N₂O), a greenhouse gas that is about 300 times more potent than carbon dioxide. It also emits nitrogen in the form of nitrogen oxides (NO and NO_2, referred to as NO_x), and unburnt ammonia may slip out, which eventually forms fine particulate matter in the atmosphere. These tiny particles can be inhaled deep into the lungs, causing health problems like heart attacks, strokes, and asthma.

The new study indicates that, under current legislation, switching the global fleet to ammonia fuel could cause up to about 600,000 additional premature deaths each year. However, with stronger regulations and cleaner engine technology, the switch could lead to about 66,000 fewer premature deaths than currently caused by maritime shipping emissions, with far less impact on global warming.

“Not all climate solutions are created equal. There is almost always some price to pay. We have to take a more holistic approach and consider all the costs and benefits of different climate solutions, rather than just their potential to decarbonize,” says Anthony Wong, a postdoc in the MIT Center for Global Change Science and lead author of the study.

His co-authors include Noelle Selin, an MIT professor in the Institute for Data, Systems, and Society and the Department of Earth, Atmospheric and Planetary Sciences (EAPS); Sebastian Eastham, a former principal research scientist who is now a senior lecturer at Imperial College London; Christine Mounaïm-Rouselle, a professor at the University of Orléans in France; Yiqi Zhang, a researcher at the Hong Kong University of Science and Technology; and Florian Allroggen, a research scientist in the MIT Department of Aeronautics and Astronautics. The research appears this week in Environmental Research Letters.

Greener, cleaner ammonia

Traditionally, ammonia is made by stripping hydrogen from natural gas and then combining it with nitrogen at extremely high temperatures. This process is often associated with a large carbon footprint. The maritime shipping industry is betting on the development of “green ammonia,” which is produced by using renewable energy to make hydrogen via electrolysis and to generate heat.

“In theory, if you are burning green ammonia in a ship engine, the carbon emissions are almost zero,” Wong says.

But even the greenest ammonia generates nitrous oxide (N₂O), nitrogen oxides (NO_x) when combusted, and some of the ammonia may slip out, unburnt. This nitrous oxide would escape into the atmosphere, where the greenhouse gas would remain for more than 100 years. At the same time, the nitrogen emitted as NO_x and ammonia would fall to Earth, damaging fragile ecosystems. As these emissions are digested by bacteria, additional N₂O  is produced.

NO_x and ammonia also mix with gases in the air to form fine particulate matter. A primary contributor to air pollution, fine particulate matter kills an estimated 4 million people each year.

“Saying that ammonia is a ‘clean’ fuel is a bit of an overstretch. Just because it is carbon-free doesn’t necessarily mean it is clean and good for public health,” Wong says.

A multifaceted model

The researchers wanted to paint the whole picture, capturing the environmental and public health impacts of switching the global fleet to ammonia fuel. To do so, they designed scenarios to measure how pollutant impacts change under certain technology and policy assumptions.

From a technological point of view, they considered two ship engines. The first burns pure ammonia, which generates higher levels of unburnt ammonia but emits fewer nitrogen oxides. The second engine technology involves mixing ammonia with hydrogen to improve combustion and optimize the performance of a catalytic converter, which controls both nitrogen oxides and unburnt ammonia pollution.

They also considered three policy scenarios: current regulations, which only limit NO_x emissions in some parts of the world; a scenario that adds ammonia emission limits over North America and Western Europe; and a scenario that adds global limits on ammonia and NO_x emissions.

The researchers used a ship track model to calculate how pollutant emissions change under each scenario and then fed the results into an air quality model. The air quality model calculates the impact of ship emissions on particulate matter and ozone pollution. Finally, they estimated the effects on global public health.

One of the biggest challenges came from a lack of real-world data, since no ammonia-powered ships are yet sailing the seas. Instead, the researchers relied on experimental ammonia combustion data from collaborators to build their model.

“We had to come up with some clever ways to make that data useful and informative to both the technology and regulatory situations,” he says.

A range of outcomes

In the end, they found that with no new regulations and ship engines that burn pure ammonia, switching the entire fleet would cause 681,000 additional premature deaths each year.

“While a scenario with no new regulations is not very realistic, it serves as a good warning of how dangerous ammonia emissions could be. And unlike NO_x, ammonia emissions from shipping are currently unregulated,” Wong says.

However, even without new regulations, using cleaner engine technology would cut the number of premature deaths down to about 80,000, which is about 20,000 fewer than are currently attributed to maritime shipping emissions. With stronger global regulations and cleaner engine technology, the number of people killed by air pollution from shipping could be reduced by about 66,000.

“The results of this study show the importance of developing policies alongside new technologies,” Selin says. “There is a potential for ammonia in shipping to be beneficial for both climate and air quality, but that requires that regulations be designed to address the entire range of potential impacts, including both climate and air quality.”

Ammonia’s air quality impacts would not be felt uniformly across the globe, and addressing them fully would require coordinated strategies across very different contexts. Most premature deaths would occur in East Asia, since air quality regulations are less stringent in this region. Higher levels of existing air pollution cause the formation of more particulate matter from ammonia emissions. In addition, shipping volume over East Asia is far greater than elsewhere on Earth, compounding these negative effects.

In the future, the researchers want to continue refining their analysis. They hope to use these findings as a starting point to urge the marine industry to share engine data they can use to better evaluate air quality and climate impacts. They also hope to inform policymakers about the importance and urgency of updating shipping emission regulations.

This research was funded by the MIT Climate and Sustainability Consortium.

While many studies over the past several years have examined people’s access to and attitudes toward Covid-19 vaccines, few studies in sub-Saharan Africa have looked at whether there were differences in vaccination rates and intention between men and women. In a new study appearing in the journal Frontiers in Global Women’s Health, researchers found that while women and men self-reported similar Covid-19 vaccination rates in 2022, unvaccinated men expressed more intention to get vaccinated than unvaccinated women.

Women tend to have better health-seeking behaviors than men overall. However, most studies relating to Covid-19 vaccination have found that intention has been lower among women. “We wondered whether this would hold true at the uptake level,” says Rawlance Ndejjo, a leader of the new study and an assistant lecturer in the Department of Disease Control and Environmental Health at Makerere University.

The comparable vaccination rates between men and women in the study is “a good thing to see,” adds Lula Chen, research director at MIT Governance Lab (GOV/LAB) and a co-author of the new study. “There wasn’t anything gendered about how [the vaccine] was being advertised or who was actually getting access to it.”

Women’s lower intention to vaccinate seemed to be driven by concerns about vaccine safety, suggesting that providing factual information about vaccine safety from trusted sources, like the Ministry of Health, could increase uptake.

The work is a collaboration between scholars from the MIT GOV/LAB, Makerere University’s School of Public Health in Uganda, University of Kinshasa’s School of Public Health in the Democratic Republic of the Congo (DRC), University of Ibadan’s College of Medicine in Nigeria, and Cheikh Anta Diop University in Senegal. 

Studying vaccine availability and uptake in sub-Saharan Africa

The authors’ collaboration began in 2021 with research into Covid-19 vaccination rates, people’s willingness to get vaccinated, and how people’s trust in different authorities shaped attitudes toward vaccines in Uganda, the DRC, Senegal, and Nigeria. A survey in Uganda found that people who received information about Covid-19 from health workers were more likely to be vaccinated, stressing the important role people who work in the health-care system can play in vaccination efforts.

Work from other scientists has found that women were less likely to accept Covid-19 vaccines than men, and that in low- and middle-income countries, women also may be less likely to get vaccinated against Covid-19 and less likely to intend to get vaccinated, possibly due to factors including lower levels of education, work obligations, and domestic care obligations.

Previous studies in sub-Saharan Africa that focused on differences between men and women with intention and willingness to vaccinate were inconclusive, Ndejjo says. “You would hardly find actual studies on uptake of the vaccines,” he adds. For the new paper, the researchers aimed to dig into uptake.

People who trust the government and health officials were more likely to get vaccinated

The researchers relied on phone survey data collected from adults in the four countries between March and July 2022. The surveys asked people about whether they’d been vaccinated and whether those who were unvaccinated intended to get vaccinated, as well as their attitudes toward Covid-19, their trust in different authorities, demographic information, and more.

Overall, 48.5 percent of men said they had been vaccinated, compared to 47.9 percent of women. Trust in authorities seemed to play a role in people’s decision to vaccinate — receiving information from health workers about Covid-19 and higher trust in the Ministry of Health were both correlated with getting vaccinated for men, whereas higher trust in the government was correlated with vaccine uptake in women.

Lower interest in vaccines among women seemed related to safety concerns

A smaller percentage of unvaccinated women (54 percent) said they intended to get vaccinated, compared to 63.4 percent of men. More unvaccinated women said they had concerns about the vaccine’s safety than unvaccinated men, which could be driving their lower intention.

The researchers also found that unvaccinated women and men over 40 had similar levels of intention to get vaccinated — lower intention in women under 40 may have driven the difference between men and women. Younger women could have concerns about vaccines related to pregnancy, Chen says. If this is the case, the research suggests that officials need to provide additional reassurance to pregnant people about vaccine safety, she adds.

Trust in authorities also contributed to people’s intention to vaccinate. Trust in the Ministry of Health was tied to higher intention to vaccinate for both men and women. Men with more trust in the World Health Organization were also more likely to intend to vaccinate.

“There’s a need to deal with a lot of the myths and misconceptions that exist,” Ndejjo says, as well as ensure that people’s concerns related to vaccine safety and effectiveness are addressed. Officials need “to work with trusted sources of information to bridge some of the gaps that we observe,” he adds. People need to be supported in their decision-making so they can make the best decisions for their health.

“This research highlights linkages between citizen trust in government, their willingness to get vaccines, and, importantly, the differences between men and women on this issue — differences that policymakers will need to understand in order to design more targeted, gender-specific public health interventions,” says study co-author Lily L. Tsai, who is MIT GOV/LAB’s director and founder and the Ford Professor of Political Science at MIT.

This project was funded by the Bill & Melinda Gates Foundation.

Researchers from the Singapore-MIT Alliance for Research and Technology (SMART), MIT’s research enterprise in Singapore, have developed a novel way to produce clinical doses of viable autologous chimeric antigen receptor (CAR) T-cells in a ultra-small automated closed-system microfluidic chip, roughly the size of a pack of cards. 

This is the first time that a microbioreactor is used to produce autologous cell therapy products. Specifically, the new method was successfully used to manufacture and expand CAR-T cells that are as effective as cells produced using existing systems in a smaller footprint and less space, and using fewer seeding cell numbers and cell manufacturing reagents. This could lead to more efficient and affordable methods of scaling-out autologous cell therapy manufacturing, and could even potentially enable point-of-care manufacturing of CAR T-cells outside of a laboratory setting — such as in hospitals and wards.

CAR T-cell therapy manufacturing requires the isolation, activation, genetic modification, and expansion of a patient’s own T-cells to kill tumor cells upon reinfusion into the patient. Despite how cell therapies have revolutionized cancer immunotherapy, with some of the first patients who received autologous cell therapies in remission for more than 10 years, the manufacturing process for CAR-T cells has remained inconsistent, costly, and time-consuming. It can be prone to contamination, subject to human error, and requires seeding cell numbers that are impractical for smaller-scale CAR T-cell production. These challenges create bottlenecks that restrict both the availability and affordability of these therapies despite their effectiveness.

In a paper titled “A high-density microbioreactor process designed for automated point-of-care manufacturing of CAR T cells” published in the journal Nature Biomedical Engineering, SMART researchers detailed their breakthrough: Human primary T-cells can be activated, transduced, and expanded to high densities in a 2-mililiter automated closed-system microfluidic chip to produce over 60 million CAR T-cells from donors with lymphoma, and over 200 million CAR T-cells from healthy donors. The CAR T-cells produced using the microbioreactor are as effective as those produced using conventional methods, but in a smaller footprint and less space, and with fewer resources. This translates to lower cost of goods manufactured (COGM), and potentially to lower costs for patients.

The groundbreaking research was led by members of the Critical Analytics for Manufacturing Personalized-Medicine (CAMP) interdisciplinary research group at SMART. Collaborators include researchers from the Duke-NUS Medical School; the Institute of Molecular and Cell Biology at the Agency for Science, Technology and Research; KK Women’s and Children’s Hospital; and Singapore General Hospital.

“This advancement in cell therapy manufacturing could ultimately offer a point-of-care platform that could substantially increase the number of CAR T-cell production slots, reducing the wait times and cost of goods of these living medicines — making cell therapy more accessible to the masses. The use of scaled-down bioreactors could also aid process optimization studies, including for different cell therapy products,” says Michael Birnbaum, co-lead principal investigator at SMART CAMP, associate professor of biological engineering at MIT, and a co-senior author of the paper.

With high T-cell expansion rates, similar total T-cell numbers could be attained with a shorter culture period in the microbioreactor (seven to eight days) compared to gas-permeable culture plates (12 days), potentially shortening production times by 30-40 percent. The CAR T-cells from both the microfluidic bioreactor and gas-permeable culture plates only showed subtle differences in cell quality. The cells were equally functional in killing leukemia cells when tested in mice.

“This new method suggests that a dramatic miniaturization of current-generation autologous cell therapy production is feasible, with the potential of significantly alleviating manufacturing limitations of CAR T-cell therapy. Such a miniaturization would lay the foundation for point-of-care manufacturing of CAR T-cells and decrease the “good manufacturing practice” (GMP) footprint required for producing cell therapies — which is one of the primary drivers of COGM,” says Wei-Xiang Sin, research scientist at SMART CAMP and first author of the paper.

Notably, the microbioreactor used in the research is a perfusion-based, automated, closed system with the smallest footprint per dose, smallest culture volume and seeding cell number, as well as the highest cell density and level of process control attainable. These microbioreactors — previously only used for microbial and mammalian cell cultures — were originally developed at MIT and have been advanced to commercial production by Millipore Sigma.

The small starting cell numbers required, compared to existing larger automated manufacturing platforms, means that smaller amounts of isolation beads, activation reagents, and lentiviral vectors are required per production run. In addition, smaller volumes of medium are required (at least tenfold lower than larger automated culture systems) owing to the extremely small culture volume (2 milliliters; approximately 100-fold lower than larger automated culture systems) — which contributes to significant reductions in reagent cost. This could benefit patients, especially pediatric patients who have low or insufficient T-cell numbers to produce therapeutic doses of CAR T-cells.

Moving forward, SMART CAMP is working on further engineering sampling and/or analytical systems around the microbioreactor so that CAR-T production can be performed with reduced labor and out of a laboratory setting, potentially facilitating the decentralized bedside manufacturing of CAR T-cells. SMART CAMP is also looking to further optimize the process parameters and culture conditions to improve cell yield and quality for future clinical use.

The research was conducted by SMART and supported by the National Research Foundation Singapore under its Campus for Research Excellence and Technological Enterprise (CREATE) program.

Some people, especially those in public service, perform admirable feats: Think of health-care workers fighting to keep patients alive or first responders arriving at the scene of a car crash. But the emotional weight can become a mental burden. Research has shown that emergency personnel are at elevated risk for mental health challenges like post-traumatic stress disorder. How can people undergo such stressful experiences and also maintain their well-being?

A new study from the McGovern Institute for Brain Research at MIT revealed that a cognitive strategy focused on social good may be effective in helping people cope with distressing events. The research team found that the approach was comparable to another well-established emotion regulation strategy, unlocking a new tool for dealing with highly adverse situations.

“How you think can improve how you feel,” says John Gabrieli, the Grover Hermann Professor of Health Sciences and Technology and a professor of brain and cognitive sciences at MIT, who is a senior author of the paper. “This research suggests that the social good approach might be particularly useful in improving well-being for those constantly exposed to emotionally taxing events.”

The study, published today in PLOS ONE, is the first to examine the efficacy of this cognitive strategy. Nancy Tsai, a postdoc in Gabrieli’s lab at the McGovern Institute, is the lead author of the paper.

Emotion regulation tools

Emotion regulation is the ability to mentally reframe how we experience emotions — a skill critical to maintaining good mental health. Doing so can make one feel better when dealing with adverse events, and emotion regulation has been shown to boost emotional, social, cognitive, and physiological outcomes across the lifespan.

One emotion regulation strategy is “distancing,” where a person copes with a negative event by imagining it as happening far away, a long time ago, or from a third-person perspective. Distancing has been well-documented as a useful cognitive tool, but it may be less effective in certain situations, especially ones that are socially charged — like a firefighter rescuing a family from a burning home. Rather than distancing themselves, a person may instead be forced to engage directly with the situation.

“In these cases, the ‘social good’ approach may be a powerful alternative,” says Tsai. “When a person uses the social good method, they view a negative situation as an opportunity to help others or prevent further harm.” For example, a firefighter experiencing emotional distress might focus on the fact that their work enables them to save lives. The idea had yet to be backed by scientific investigation, so Tsai and her team, alongside Gabrieli, saw an opportunity to rigorously probe this strategy.

A novel study

The MIT researchers recruited a cohort of adults and had them complete a questionnaire to gather information including demographics, personality traits, and current well-being, as well as how they regulated their emotions and dealt with stress. The cohort was randomly split into two groups: a distancing group and a social good group. In the online study, each group was shown a series of images that were either neutral (such as fruit) or contained highly aversive content (such as bodily injury). Participants were fully informed of the kinds of images they might see and could opt out of the study at any time.

Each group was asked to use their assigned cognitive strategy to respond to half of the negative images. For example, while looking at a distressing image, a person in the distancing group could have imagined that it was a screenshot from a movie. Conversely, a subject in the social good group might have responded to the image by envisioning that they were a first responder saving people from harm. For the other half of the negative images, participants were asked to only look at them and pay close attention to their emotions. The researchers asked the participants how they felt after each image was shown.

Social good as a potent strategy

The MIT team found that distancing and social good approaches helped diminish negative emotions. Participants reported feeling better when they used these strategies after viewing adverse content compared to when they did not, and stated that both strategies were easy to implement.

The results also revealed that, overall, distancing yielded a stronger effect. Importantly, however, Tsai and Gabrieli believe that this study offers compelling evidence for social good as a powerful method better-suited to situations when people cannot distance themselves, like rescuing someone from a car crash, “Which is more probable for people in the real world,” notes Tsai. Moreover, the team discovered that people who most successfully used the social good approach were more likely to view stress as enhancing rather than debilitating. Tsai says this link may point to psychological mechanisms that underlie both emotion regulation and how people respond to stress.

Additionally, the results showed that older adults used the cognitive strategies more effectively than younger adults. The team suspects that this is probably because, as prior research has shown, older adults are more adept at regulating their emotions, likely due to having greater life experiences. The authors note that successful emotion regulation also requires cognitive flexibility, or having a malleable mindset to adapt well to different situations.

“This is not to say that people, such as physicians, should reframe their emotions to the point where they fully detach themselves from negative situations,” says Gabrieli. “But our study shows that the social good approach may be a potent strategy to combat the immense emotional demands of certain professions.”

The MIT team says that future studies are needed to further validate this work, and that such research is promising in that it can uncover new cognitive tools to equip individuals to take care of themselves as they bravely assume the challenge of taking care of others.

As climate change advances, the ocean’s overturning circulation is predicted to weaken substantially. With such a slowdown, scientists estimate the ocean will pull down less carbon dioxide from the atmosphere. However, a slower circulation should also dredge up less carbon from the deep ocean that would otherwise be released back into the atmosphere. On balance, the ocean should maintain its role in reducing carbon emissions from the atmosphere, if at a slower pace.

However, a new study by an MIT researcher finds that scientists may have to rethink the relationship between the ocean’s circulation and its long-term capacity to store carbon. As the ocean gets weaker, it could release more carbon from the deep ocean into the atmosphere instead.

The reason has to do with a previously uncharacterized feedback between the ocean’s available iron, upwelling carbon and nutrients, surface microorganisms, and a little-known class of molecules known generally as “ligands.” When the ocean circulates more slowly, all these players interact in a self-perpetuating cycle that ultimately increases the amount of carbon that the ocean outgases back to the atmosphere.

“By isolating the impact of this feedback, we see a fundamentally different relationship between ocean circulation and atmospheric carbon levels, with implications for the climate,” says study author Jonathan Lauderdale, a research scientist in MIT’s Department of Earth, Atmospheric, and Planetary Sciences. “What we thought is going on in the ocean is completely overturned.”

Lauderdale says the findings show that “we can’t count on the ocean to store carbon in the deep ocean in response to future changes in circulation. We must be proactive in cutting emissions now, rather than relying on these natural processes to buy us time to mitigate climate change.”

His study appears today in the journal Nature Communications.

Box flow

In 2020, Lauderdale led a study that explored ocean nutrients, marine organisms, and iron, and how their interactions influence the growth of phytoplankton around the world. Phytoplankton are microscopic, plant-like organisms that live on the ocean surface and consume a diet of carbon and nutrients that upwell from the deep ocean and iron that drifts in from desert dust.

The more phytoplankton that can grow, the more carbon dioxide they can absorb from the atmosphere via photosynthesis, and this plays a large role in the ocean’s ability to sequester carbon.

For the 2020 study, the team developed a simple “box” model, representing conditions in different parts of the ocean as general boxes, each with a different balance of nutrients, iron, and ligands — organic molecules that are thought to be byproducts of phytoplankton. The team modeled a general flow between the boxes to represent the ocean’s larger circulation — the way seawater sinks, then is buoyed back up to the surface in different parts of the world.

This modeling revealed that, even if scientists were to “seed” the oceans with extra iron, that iron wouldn’t have much of an effect on global phytoplankton growth. The reason was due to a limit set by ligands. It turns out that, if left on its own, iron is insoluble in the ocean and therefore unavailable to phytoplankton. Iron only becomes soluble at “useful” levels when linked with ligands, which keep iron in a form that plankton can consume. Lauderdale found that adding iron to one ocean region to consume additional nutrients robs other regions of nutrients that phytoplankton there need to grow. This lowers the production of ligands and the supply of iron back to the original ocean region, limiting the amount of extra carbon that would be taken up from the atmosphere.

Unexpected switch

Once the team published their study, Lauderdale worked the box model into a form that he could make publicly accessible, including ocean and atmosphere carbon exchange and extending the boxes to represent more diverse environments, such as conditions similar to the Pacific, the North Atlantic, and the Southern Ocean. In the process, he tested other interactions within the model, including the effect of varying ocean circulation.

He ran the model with different circulation strengths, expecting to see less atmospheric carbon dioxide with weaker ocean overturning — a relationship that previous studies have supported, dating back to the 1980s. But what he found instead was a clear and opposite trend: The weaker the ocean’s circulation, the more CO₂ built up in the atmosphere.

“I thought there was some mistake,” Lauderdale recalls. “Why were atmospheric carbon levels trending the wrong way?”

When he checked the model, he found that the parameter describing ocean ligands had been left “on” as a variable. In other words, the model was calculating ligand concentrations as changing from one ocean region to another.

On a hunch, Lauderdale turned this parameter “off,” which set ligand concentrations as constant in every modeled ocean environment, an assumption that many ocean models typically make. That one change reversed the trend, back to the assumed relationship: A weaker circulation led to reduced atmospheric carbon dioxide. But which trend was closer to the truth?

Lauderdale looked to the scant available data on ocean ligands to see whether their concentrations were more constant or variable in the actual ocean. He found confirmation in GEOTRACES, an international study that coordinates measurements of trace elements and isotopes across the world’s oceans, that scientists can use to compare concentrations from region to region. Indeed, the molecules’ concentrations varied. If ligand concentrations do change from one region to another, then his surprise new result was likely representative of the real ocean: A weaker circulation leads to more carbon dioxide in the atmosphere.

“It’s this one weird trick that changed everything,” Lauderdale says. “The ligand switch has revealed this completely different relationship between ocean circulation and atmospheric CO₂ that we thought we understood pretty well.”

Slow cycle

To see what might explain the overturned trend, Lauderdale analyzed biological activity and carbon, nutrient, iron, and ligand concentrations from the ocean model under different circulation strengths, comparing scenarios where ligands were variable or constant across the various boxes.

This revealed a new feedback: The weaker the ocean’s circulation, the less carbon and nutrients the ocean pulls up from the deep. Any phytoplankton at the surface would then have fewer resources to grow and would produce fewer byproducts (including ligands) as a result. With fewer ligands available, less iron at the surface would be usable, further reducing the phytoplankton population. There would then be fewer phytoplankton available to absorb carbon dioxide from the atmosphere and consume upwelled carbon from the deep ocean.

“My work shows that we need to look more carefully at how ocean biology can affect the climate,” Lauderdale points out. “Some climate models predict a 30 percent slowdown in the ocean circulation due to melting ice sheets, particularly around Antarctica. This huge slowdown in overturning circulation could actually be a big problem: In addition to a host of other climate issues, not only would the ocean take up less anthropogenic CO₂ from the atmosphere, but that could be amplified by a net outgassing of deep ocean carbon, leading to an unanticipated increase in atmospheric CO₂ and unexpected further climate warming.” 

A new tool makes it easier for database users to perform complicated statistical analyses of tabular data without the need to know what is going on behind the scenes.

GenSQL, a generative AI system for databases, could help users make predictions, detect anomalies, guess missing values, fix errors, or generate synthetic data with just a few keystrokes.

For instance, if the system were used to analyze medical data from a patient who has always had high blood pressure, it could catch a blood pressure reading that is low for that particular patient but would otherwise be in the normal range.

GenSQL automatically integrates a tabular dataset and a generative probabilistic AI model, which can account for uncertainty and adjust their decision-making based on new data.

Moreover, GenSQL can be used to produce and analyze synthetic data that mimic the real data in a database. This could be especially useful in situations where sensitive data cannot be shared, such as patient health records, or when real data are sparse.

This new tool is built on top of SQL, a programming language for database creation and manipulation that was introduced in the late 1970s and is used by millions of developers worldwide.

“Historically, SQL taught the business world what a computer could do. They didn’t have to write custom programs, they just had to ask questions of a database in high-level language. We think that, when we move from just querying data to asking questions of models and data, we are going to need an analogous language that teaches people the coherent questions you can ask a computer that has a probabilistic model of the data,” says Vikash Mansinghka ’05, MEng ’09, PhD ’09, senior author of a paper introducing GenSQL and a principal research scientist and leader of the Probabilistic Computing Project in the MIT Department of Brain and Cognitive Sciences.

When the researchers compared GenSQL to popular, AI-based approaches for data analysis, they found that it was not only faster but also produced more accurate results. Importantly, the probabilistic models used by GenSQL are explainable, so users can read and edit them.

“Looking at the data and trying to find some meaningful patterns by just using some simple statistical rules might miss important interactions. You really want to capture the correlations and the dependencies of the variables, which can be quite complicated, in a model. With GenSQL, we want to enable a large set of users to query their data and their model without having to know all the details,” adds lead author Mathieu Huot, a research scientist in the Department of Brain and Cognitive Sciences and member of the Probabilistic Computing Project.

They are joined on the paper by Matin Ghavami and Alexander Lew, MIT graduate students; Cameron Freer, a research scientist; Ulrich Schaechtle and Zane Shelby of Digital Garage; Martin Rinard, an MIT professor in the Department of Electrical Engineering and Computer Science and member of the Computer Science and Artificial Intelligence Laboratory (CSAIL); and Feras Saad ’15, MEng ’16, PhD ’22, an assistant professor at Carnegie Mellon University. The research was recently presented at the ACM Conference on Programming Language Design and Implementation.

Combining models and databases

SQL, which stands for structured query language, is a programming language for storing and manipulating information in a database. In SQL, people can ask questions about data using keywords, such as by summing, filtering, or grouping database records.

However, querying a model can provide deeper insights, since models can capture what data imply for an individual. For instance, a female developer who wonders if she is underpaid is likely more interested in what salary data mean for her individually than in trends from database records.

The researchers noticed that SQL didn’t provide an effective way to incorporate probabilistic AI models, but at the same time, approaches that use probabilistic models to make inferences didn’t support complex database queries.

They built GenSQL to fill this gap, enabling someone to query both a dataset and a probabilistic model using a straightforward yet powerful formal programming language.

A GenSQL user uploads their data and probabilistic model, which the system automatically integrates. Then, she can run queries on data that also get input from the probabilistic model running behind the scenes. This not only enables more complex queries but can also provide more accurate answers.

For instance, a query in GenSQL might be something like, “How likely is it that a developer from Seattle knows the programming language Rust?” Just looking at a correlation between columns in a database might miss subtle dependencies. Incorporating a probabilistic model can capture more complex interactions.   

Plus, the probabilistic models GenSQL utilizes are auditable, so people can see which data the model uses for decision-making. In addition, these models provide measures of calibrated uncertainty along with each answer.

For instance, with this calibrated uncertainty, if one queries the model for predicted outcomes of different cancer treatments for a patient from a minority group that is underrepresented in the dataset, GenSQL would tell the user that it is uncertain, and how uncertain it is, rather than overconfidently advocating for the wrong treatment.

Faster and more accurate results

To evaluate GenSQL, the researchers compared their system to popular baseline methods that use neural networks. GenSQL was between 1.7 and 6.8 times faster than these approaches, executing most queries in a few milliseconds while providing more accurate results.

They also applied GenSQL in two case studies: one in which the system identified mislabeled clinical trial data and the other in which it generated accurate synthetic data that captured complex relationships in genomics.

Next, the researchers want to apply GenSQL more broadly to conduct largescale modeling of human populations. With GenSQL, they can generate synthetic data to draw inferences about things like health and salary while controlling what information is used in the analysis.

They also want to make GenSQL easier to use and more powerful by adding new optimizations and automation to the system. In the long run, the researchers want to enable users to make natural language queries in GenSQL. Their goal is to eventually develop a ChatGPT-like AI expert one could talk to about any database, which grounds its answers using GenSQL queries.   

This research is funded, in part, by the Defense Advanced Research Projects Agency (DARPA), Google, and the Siegel Family Foundation.

Microbes that are used for health, agricultural, or other applications need to be able to withstand extreme conditions, and ideally the manufacturing processes used to make tablets for long-term storage. MIT researchers have now developed a new way to make microbes hardy enough to withstand these extreme conditions.

Their method involves mixing bacteria with food and drug additives from a list of compounds that the FDA classifies as “generally regarded as safe.” The researchers identified formulations that help to stabilize several different types of microbes, including yeast and bacteria, and they showed that these formulations could withstand high temperatures, radiation, and industrial processing that can damage unprotected microbes.

In an even more extreme test, some of the microbes recently returned from a trip to the International Space Station, coordinated by Space Center Houston Manager of Science and Research Phyllis Friello, and the researchers are now analyzing how well the microbes were able to withstand those conditions.

“What this project was about is stabilizing organisms for extreme conditions. We're thinking about a broad set of applications, whether it's missions to space, human applications, or agricultural uses,” says Giovanni Traverso, an associate professor of mechanical engineering at MIT, a gastroenterologist at Brigham and Women’s Hospital, and the senior author of the study.

Miguel Jimenez, a former MIT research scientist who is now an assistant professor of biomedical engineering at Boston University, is the lead author of the paper, which appears today in Nature Materials.

Surviving extreme conditions

About six years ago, with funding from NASA’s Translational Research Institute for Space Health (TRISH), Traverso’s lab began working on new approaches to make helpful bacteria such as probiotics and microbial therapeutics more resilient. As a starting point, the researchers analyzed 13 commercially available probiotics and found that six of these products did not contain as many live bacteria as the label indicated.

“What we found was that, perhaps not surprisingly, there is a difference, and it can be significant,” Traverso says. “So then the next question was, given this, what can we do to help the situation?”

For their experiments, the researchers chose four different microbes to focus on: three bacteria and one yeast. These microbes are Escherichia coli Nissle 1917, a probiotic; Ensifer meliloti, a bacterium that can fix nitrogen in soil to support plant growth; Lactobacillus plantarum, a bacterium used to ferment food products; and the yeast Saccharomyces boulardii, which is also used as a probiotic.

When microbes are used for medical or agricultural applications, they are usually dried into a powder through a process called lyophilization. However, they can not normally be made into more useful forms such as a tablet or pill because this process requires exposure to an organic solvent, which can be toxic to the bacteria. The MIT team set out to find additives that could improve the microbes’ ability to survive this kind of processing.

“We developed a workflow where we can take materials from the ‘generally regarded as safe’ materials list from the FDA, and mix and match those with bacteria and ask, are there ingredients that enhance the stability of the bacteria during the lyophilization process?” Traverso says.

Their setup allows them to mix microbes with one of about 100 different ingredients and then grow them to see which survive the best when stored at room temperature for 30 days. These experiments revealed different ingredients, mostly sugars and peptides, that worked best for each species of microbe.

The researchers then picked one of the microbes, E. coli Nissle 1917, for further optimization. This probiotic has been used to treat “traveler’s diarrhea,” a condition caused by drinking water contaminated with harmful bacteria. The researchers found that if they combined caffeine or yeast extract with a sugar called melibiose, they could create a very stable formulation of E. coli Nissle 1917. This mixture, which the researchers called formulation D, allowed survival rates greater than 10 percent after the microbes were stored for six months at 37 degrees Celsius, while a commercially available formulation of E. coli Nissle 1917 lost all viability after only 11 days under those conditions.

Formulation D was also able to withstand much higher levels of ionizing radiation, up to 1,000 grays. (The typical radiation dose on Earth is about 15 micrograys per day, and in space, it’s about 200 micrograys per day.)

The researchers don’t know exactly how their formulations protect bacteria, but they hypothesize that the additives may help to stabilize the bacterial cell membranes during rehydration.

Stress tests

The researchers then showed that these microbes can not only survive harsh conditions, they also maintain their function after these exposures. After Ensifer meliloti were exposed to temperatures up to 50 degrees Celsius, the researchers found that they were still able to form symbiotic nodules on plant roots and convert nitrogen to ammonia.

They also found that their formulation of E. coli Nissle 1917 was able to inhibit the growth of Shigellaflexneri, one of the leading causes of diarrhea-associated deaths in low- and middle-income countries, when the microbes were grown together in a lab dish.

Last year, several strains of these extremophile microbes were sent to the International Space Station, which Jimenez describes as “the ultimate stress test.”

“Even just the shipping on Earth to the preflight validation, and storage until flight are part of this test, with no temperature control along the way,” he says.

The samples recently returned to Earth, and Jimenez’ lab is now analyzing them. He plans to compare samples that were kept inside the ISS to others that were bolted to the outside of the station, as well as control samples that remained on Earth.

“This work offers a promising approach to enhance the stability of probiotics and/or genetically engineered  microbes in extreme environments, such as in outer space, which could be used in future space missions to help maintain astronaut health or promote sustainability, such as in promoting more robust and resilient plants for food production,” says Camilla Urbaniak, a research scientist at NASA’s Jet Propulsion Laboratory, who was not involved in the study.

The research was funded by NASA’s Translational Research Institute for Space Health, Space Center Houston, MIT’s Department of Mechanical Engineering, and by 711 Human Performance Wing and the Defense Advanced Research Projects Agency.

Other authors of the paper include Johanna L’Heureux, Emily Kolaya, Gary Liu, Kyle Martin, Husna Ellis, Alfred Dao, Margaret Yang, Zachary Villaverde, Afeefah Khazi-Syed, Qinhao Cao, Niora Fabian, Joshua Jenkins, Nina Fitzgerald, Christina Karavasili, Benjamin Muller, and James Byrne.

While recycling systems and bottle deposits have become increasingly widespread in the U.S., actual rates of recycling are “abysmal,” according to a team of MIT researchers who studied the rates for recycling of PET, the plastic commonly used in beverage bottles. However, their findings suggest some ways to change this.

The present rate of recycling for PET, or polyethylene terephthalate, bottles nationwide is about 24 percent and has remained stagnant for a decade, the researchers say. But their study indicates that with a nationwide bottle deposit program, the rates could increase to 82 percent, with nearly two-thirds of all PET bottles being recycled into new bottles, at a net cost of just a penny a bottle when demand is robust. At the same time, they say, policies would be needed to ensure a sufficient demand for the recycled material.

The findings are being published today in the Journal of Industrial Ecology, in a paper by MIT professor of materials science and engineering Elsa Olivetti, graduate students Basuhi Ravi and Karan Bhuwalka, and research scientist Richard Roth.

The team looked at PET bottle collection and recycling rates in different states as well as other nations with and without bottle deposit policies, and with or without curbside recycling programs, as well as the inputs and outputs of various recycling companies and methods. The researchers say this study is the first to look in detail at the interplay between public policies and the end-to-end realities of the packaging production and recycling market.

They found that bottle deposit programs are highly effective in the areas where they are in place, but at present there is not nearly enough collection of used bottles to meet the targets set by the packaging industry. Their analysis suggests that a uniform nationwide bottle deposit policy could achieve the levels of recycling that have been mandated by proposed legislation and corporate commitments.

The recycling of PET is highly successful in terms of quality, with new products made from all-recycled material virtually matching the qualities of virgin material. And brands have shown that new bottles can be safely made with 100 percent postconsumer waste. But the team found that collection of the material is a crucial bottleneck that leaves processing plants unable to meet their needs. However, with the right policies in place, “one can be optimistic,” says Olivetti, who is the Jerry McAfee Professor in Engineering and the associate dean of the School of Engineering.

“A message that we have found in a number of cases in the recycling space is that if you do the right work to support policies that think about both the demand but also the supply,” then significant improvements are possible, she says. “You have to think about the response and the behavior of multiple actors in the system holistically to be viable,” she says. “We are optimistic, but there are many ways to be pessimistic if we’re not thinking about that in a holistic way.”

For example, the study found that it is important to consider the needs of existing municipal waste-recovery facilities. While expanded bottle deposit programs are essential to increase recycling rates and provide the feedstock to companies recycling PET into new products, the current facilities that process material from curbside recycling programs will lose revenue from PET bottles, which are a relatively high-value product compared to the other materials in the recycled waste stream. These companies would lose a source of their income if the bottles are collected through deposit programs, leaving them with only the lower-value mixed plastics.

The researchers developed economic models based on rates of collection found in the states with deposit programs, recycled-content requirements, and other policies, and used these models to extrapolate to the nation as a whole. Overall, they found that the supply needs of packaging producers could be met through a nationwide bottle deposit system with a 10-cent deposit per bottle — at a net cost of about 1 cent per bottle produced when demand is strong. This need not be a federal program, but rather one where the implementation would be left up to the individual states, Olivetti says.

Other countries have been much more successful in implementing deposit systems that result in very high participation rates. Several European countries manage to collect more than 90 percent of PET bottles for recycling, for example. But in the U.S., less than 29 percent are collected, and after losses in the recycling chain about 24 percent actually get recycled, the researchers found. Whereas 73 percent of Americans have access to curbside recycling, presently only 10 states have bottle deposit systems in place.

Yet the demand is there so far. “There is a market for this material,” says Olivetti. While bottles collected through mixed-waste collection can still be recycled to some extent, those collected through deposit systems tend to be much cleaner and require less processing, and so are more economical to recycle into new bottles, or into textiles.

To be effective, policies need to not just focus on increasing rates of recycling, but on the whole cycle of supply and demand and the different players involved, Olivetti says. Safeguards would need to be in place to protect existing recycling facilities from the lost revenues they would suffer as a result of bottle deposits, perhaps in the form of subsidies funded by fees on the bottle producers, to avoid putting these essential parts of the processing chain out of business. And other policies may be needed to ensure the continued market for the material that gets collected, including recycled content requirements and extended producer responsibility regulations, the team found.

At this stage, it’s important to focus on the specific waste streams that can most effectively be recycled, and PET, along with many metals, clearly fit that category. “When we start to think about mixed plastic streams, that’s much more challenging from an environmental perspective,” she says. “Recycling systems need to be pursuing extended producers’ responsibility, or specifically thinking about materials designed more effectively toward recycled content,” she says.

It's also important to address “what the right metrics are to design for sustainably managed materials streams,” she says. “It could be energy use, could be circularity [for example, making old bottles into new bottles], could be around waste reduction, and making sure those are all aligned. That’s another kind of policy coordination that’s needed.”

At the core of Raymond Wang’s work lies a seemingly simple question: Can’t we just get along?

Wang, a fifth-year political science graduate student, is a native of Hong Kong who witnessed firsthand the shakeup and conflict engendered by China’s takeover of the former British colony. “That type of experience makes you wonder why things are so complicated,” he says. “Why is it so hard to live with your neighbors?”

Today, Wang is focused on ways of managing a rapidly intensifying U.S.-China competition, and more broadly, on identifying how China — and other emerging global powers — bend, break, or creatively accommodate international rules in trade, finance, maritime, and arms control matters to achieve their ends.

The current game for global dominance between the United States and China continually threatens to erupt into dangerous confrontation. Wang’s research aims to construct a more nuanced take on China’s behaviors in this game.

“U.S. policy towards China should be informed by a better understanding of China’s behaviors if we are to avoid the worst-case scenario,” Wang believes.

“Selective and smart”

One of Wang’s major research thrusts is the ongoing trade war between the two nations. “The U.S. views China as rewriting the rules, creating an alternative world order — and accuses China of violating World Trade Organization (WTO) rules,” says Wang. “But in fact, China has been very selective and smart about responding to these rules.”

One critical, and controversial, WTO matter involves determining whether state-owned enterprises are, in the arcane vocabulary of the group, “public bodies,” which are subject to sometimes punitive WTO rules. The United States asserts that if a government owns 51 percent of a company, it is a public body. This means that many essential Chinese state-owned enterprises (SOEs) — manufacturers of electric vehicles, steel, or chemicals, for example — would fall under WTO provisions, and potentially face punitive discipline.

But China isn’t the only nation with SOEs. Many European countries, including stalwart U.S. partners France and Norway, subsidize companies that qualify as public bodies according to the U.S. definition. They, too, could be subject to tough WTO regulations.

“This could harm a swathe of the E.U. economy,” says Wang. “So China intelligently made the case to the international community that the U.S. position is extreme, and has pushed for a more favorable interpretation through litigation at the WTO.”

For Wang, this example highlights a key insight of his research: “Rising powers such as China exhibit cautious opportunism,” he says. “China will try to work with the existing rules as much as possible, including bending them in creative ways.”

But when it comes down to it, Wang argues, China would rather avoid the costs of building something completely new.

“If you can repurpose an old tool, why would you buy a new one?” he asks. “The vast majority of actions China is taking involves reshaping the existing order, not introducing new rules or blowing up institutions and building new ones.”

Interviewing key players

To bolster his theory of “cautious opportunism,” Wang’s doctoral project sets out a suite of rule-shaping strategies adopted by rising powers in international organizations. His analysis is driven by case studies of disputes recently concluded, or ongoing, in the WTO, the World Bank, and other bodies responsible for defining and policing rules that govern all manner of international relations and commerce.

Gathering evidence for his argument, Wang has been interviewing people critical to the disputes on all sides.

“My approach is to figure out who was in the room when certain decisions were made and talk to every single person there,” he says. “For the WTO and World Bank, I’ve interviewed close to 50 relevant personnel, including front-line lawyers, senior leadership, and former government officials.” These interviews took place in Geneva, Singapore, Tokyo, and Washington.

But writing about disputes that involve China poses a unique set of problems. “It’s difficult to talk to actively serving Chinese officials, and in general, nobody wants to go on the record because all the content is sensitive.” 

As Wang moves on to cases in maritime governance, he will be reaching out to the key players involved in managing sensitive conflicts in the South China Sea, an Indo-Pacific region dotted with shoals and offering desirable fisheries as well as oil and gas resources.

Even here, Wang suggests, China may find reason to be cautious rather than opportunistic, preferring to carve out exemptions for itself or shift interpretations, rather than overturning the existing rules wholesale.

Indeed, Wang believes China and other rising powers introduce new rules only when conditions open up a window of opportunity: “It may be worth doing so when using traditional tools doesn’t get you what you want, if your competitors are unable or unwilling to counter mobilize against you, and you see that the costs of establishing these new rules are worth it,” he says.

Beyond Wang’s dissertation, he has also been part of a research team led by M. Taylor Fravel, Arthur and Ruth Sloan Professor of Political Science, that has published papers on China’s Belt and Road Initiative.

From friends to enemies

Wang left Hong Kong and its political ferment behind at age 15, but the challenge of dealing with a powerful neighbor and the potential crisis it represented stayed with him. In Italy, he attended a United World College — part of a network of schools bringing together young people from different nations and cultures for the purpose of training leaders and peacemakers.

“It’s a utopian idea, where you force teenagers from all around the world to live and study together and get along for two years,” says Wang. “There were people from countries in the Balkans that were actively at war with each other, who grew up with the memory of air raid sirens and family members who fought each other, but these kids would just hang out together.”  

Coexistence was possible on the individual level, Wang realized, but he wondered, “What systemic thing happens that makes people do messed-up stuff to each other when they are in a group?”

With this question in mind, he went to the University of St. Andrews for his undergraduate and master’s degrees in international relations and modern history. As China continued its economic and military march onto the world stage, and Iran generated international tensions over its nuclear ambitions, Wang became interested in nuclear disarmament. He drilled down into the subject at the Middlebury Institute of International Studies at Monterey, where he earned a second master’s degree in nonproliferation and terrorism studies.

Leaning into a career revolving around policy, he applied to MIT’s security studies doctoral program, hoping to focus on the impact of emerging technologies on strategic nuclear stability. But events in the world led him to pivot. “When I started in the fall of 2019, the U.S.-China relationship was going off the rails with the trade war,” he says. “It was clear that managing the relationship would be one of the biggest foreign policy challenges for the foreseeable future, and I wanted to do research that would help ensure that the relationship wouldn’t tip into a nuclear war.”

Cooling tensions

Wang has no illusions about the difficulty of containing tensions between a superpower eager to assert its role in the world order, and one determined to hold onto its primacy. His goal is to make the competition more transparent, and if possible, less overtly threatening. He is preparing a paper, “Guns and Butter: Measuring Spillover and Implications for Technological Competition,” that outlines the different paths taken by the United States and China in developing defense-related technology that also benefits the civilian economy.

As he wades into the final phase of his thesis and contemplates his next steps, Wang hopes that his research insights might inform policymakers, especially in the United States, in their approach to China. While there is a fiercely competitive relationship, “there is still room for diplomacy,” he believes. “If you accept my theory that a rising power will try and use, or even abuse,  existing rules as much as possible, then you need non-military — State Department — boots on the ground to monitor what is going on at all the international institutions,” he says. The more information and understanding the United States has of China’s behavior, the more likely it will be able “to cool down some of the tensions,” says Wang. “We need to develop a strategic empathy.”
 

Titanium alloys are essential structural materials for a wide variety of applications, from aerospace and energy infrastructure to biomedical equipment. But like most metals, optimizing their properties tends to involve a tradeoff between two key characteristics: strength and ductility. Stronger materials tend to be less deformable, and deformable materials tend to be mechanically weak.

Now, researchers at MIT, collaborating with researchers at ATI Specialty Materials, have discovered an approach for creating new titanium alloys that can exceed this historical tradeoff, leading to new alloys with exceptional combinations of strength and ductility, which might lead to new applications.

The findings are described in the journal Advanced Materials, in a paper by Shaolou Wei ScD ’22, Professor C. Cem Tasan, postdoc Kyung-Shik Kim, and John Foltz from ATI Inc. The improvements, the team says, arise from tailoring the chemical composition and the lattice structure of the alloy, while also adjusting the processing techniques used to produce the material at industrial scale.

Titanium alloys have been important because of their exceptional mechanical properties, corrosion resistance, and light weight when compared to steels for example. Through careful selection of the alloying elements and their relative proportions, and of the way the material is processed, “you can create various different structures, and this creates a big playground for you to get good property combinations, both for cryogenic and elevated temperatures,” Tasan says.

But that big assortment of possibilities in turn requires a way to guide the selections to produce a material that meets the specific needs of a particular application. The analysis and experimental results described in the new study provide that guidance.

The structure of titanium alloys, all the way down to atomic scale, governs their properties, Tasan explains. And in some titanium alloys, this structure is even more complex, made up of two different intermixed phases, known as the alpha and beta phases.

“The key strategy in this design approach is to take considerations of different scales,” he says. “One scale is the structure of individual crystal. For example, by choosing the alloying elements carefully, you can have a more ideal crystal structure of the alpha phase that enables particular deformation mechanisms. The other scale is the polycrystal scale, that involves interactions of the alpha and beta phases. So, the approach that’s followed here involves design considerations for both.”

In addition to choosing the right alloying materials and proportions, steps in the processing turned out to play an important role. A technique called cross-rolling is another key to achieving the exceptional combination of strength and ductility, the team found.

Working together with ATI researchers, the team tested a variety of alloys under a scanning electron microscope as they were being deformed, revealing details of how their microstructures respond to external mechanical load. They found that there was a particular set of parameters — of composition, proportions, and processing method — that yielded a structure where the alpha and beta phases shared the deformation uniformly, mitigating the cracking tendency that is likely to occur between the phases when they respond differently. “The phases deform in harmony,” Tasan says. This cooperative response to deformation can yield a superior material, they found.

“We looked at the structure of the material to understand these two phases and their morphologies, and we looked at their chemistries by carrying out local chemical analysis at the atomic scale. We adopted a wide variety of techniques to quantify various properties of the material across multiple length scales, says Tasan, who is the POSCO Professor of Materials Science and Engineering and an associate professor of metallurgy. “When we look at the overall properties” of the titanium alloys produced according to their system, “the properties are really much better than comparable alloys.”

This was industry-supported academic research aimed at proving design principles for alloys that can be commercially produced at scale, according to Tasan. “What we do in this collaboration is really toward a fundamental understanding of crystal plasticity,” he says. “We show that this design strategy is validated, and we show scientifically how it works,” he adds, noting that there remains significant room for further improvement.

As for potential applications of these findings, he says, “for any aerospace application where an improved combination of strength and ductility are useful, this kind of invention is providing new opportunities.”

The work was supported by ATI Specialty Rolled Products and used facilities of MIT.nano and the Center for Nanoscale Systems at Harvard University.

Cochlear implants, tiny electronic devices that can provide a sense of sound to people who are deaf or hard of hearing, have helped improve hearing for more than a million people worldwide, according to the National Institutes of Health.

However, cochlear implants today are only partially implanted, and they rely on external hardware that typically sits on the side of the head. These components restrict users, who can’t, for instance, swim, exercise, or sleep while wearing the external unit, and they may cause others to forgo the implant altogether.

On the way to creating a fully internal cochlear implant, a multidisciplinary team of researchers at MIT, Massachusetts Eye and Ear, Harvard Medical School, and Columbia University has produced an implantable microphone that performs as well as commercial external hearing aid microphones. The microphone remains one of the largest roadblocks to adopting a fully internalized cochlear implant.

This tiny microphone, a sensor produced from a biocompatible piezoelectric material, measures miniscule movements on the underside of the ear drum. Piezoelectric materials generate an electric charge when compressed or stretched. To maximize the device’s performance, the team also developed a low-noise amplifier that enhances the signal while minimizing noise from the electronics.

While many challenges must be overcome before such a microphone could be used with a cochlear implant, the collaborative team looks forward to further refining and testing this prototype, which builds off work begun at MIT and Mass Eye and Ear more than a decade ago.

“It starts with the ear doctors who are with this every day of the week, trying to improve people’s hearing, recognizing a need, and bringing that need to us. If it weren’t for this team collaboration, we wouldn’t be where we are today,” says Jeffrey Lang, the Vitesse Professor of Electrical Engineering, a member of the Research Laboratory of Electronics (RLE), and co-senior author of a paper on the microphone.

Lang’s coauthors include co-lead authors Emma Wawrzynek, an electrical engineering and computer science (EECS) graduate student, and Aaron Yeiser SM ’21; as well as mechanical engineering graduate student John Zhang; Lukas Graf and Christopher McHugh of Mass Eye and Ear; Ioannis Kymissis, the Kenneth Brayer Professor of Electrical Engineering at Columbia; Elizabeth S. Olson, a professor of biomedical engineering and auditory biophysics at Columbia; and co-senior author Hideko Heidi Nakajima, an associate professor of otolaryngology-head and neck surgery at Harvard Medical School and Mass Eye and Ear. The research is published today in the Journal of Micromechanics and Microengineering.

Overcoming an implant impasse

Cochlear implant microphones are usually placed on the side of the head, which means that users can’t take advantage of noise filtering and sound localization cues provided by the structure of the outer ear.

Fully implantable microphones offer many advantages. But most devices currently in development, which sense sound under the skin or motion of middle ear bones, can struggle to capture soft sounds and wide frequencies.

For the new microphone, the team targeted a part of the middle ear called the umbo. The umbo vibrates unidirectionally (inward and outward), making it easier to sense these simple movements.

Although the umbo has the largest range of movement of the middle-ear bones, it only moves by a few nanometers. Developing a device to measure such diminutive vibrations presents its own challenges.

On top of that, any implantable sensor must be biocompatible and able to withstand the body’s humid, dynamic environment without causing harm, which limits the materials that can be used.

“Our goal is that a surgeon implants this device at the same time as the cochlear implant and internalized processor, which means optimizing the surgery while working around the internal structures of the ear without disrupting any of the processes that go on in there,” Wawrzynek says.

With careful engineering, the team overcame these challenges.

They created the UmboMic, a triangular, 3-millimeter by 3-millimeter motion sensor composed of two layers of a biocompatible piezoelectric material called polyvinylidene difluoride (PVDF). These PVDF layers are sandwiched on either side of a flexible printed circuit board (PCB), forming a microphone that is about the size of a grain of rice and 200 micrometers thick. (An average human hair is about 100 micrometers thick.)

The narrow tip of the UmboMic would be placed against the umbo. When the umbo vibrates and pushes against the piezoelectric material, the PVDF layers bend and generate electric charges, which are measured by electrodes in the PCB layer.

Amplifying performance

The team used a “PVDF sandwich” design to reduce noise. When the sensor is bent, one layer of PVDF produces a positive charge and the other produces a negative charge. Electrical interference adds to both equally, so taking the difference between the charges cancels out the noise.

Using PVDF provides many advantages, but the material made fabrication especially difficult. PVDF loses its piezoelectric properties when exposed to temperatures above around 80 degrees Celsius, yet very high temperatures are needed to vaporize and deposit titanium, another biocompatible material, onto the sensor. Wawrzynek worked around this problem by depositing the titanium gradually and employing a heat sink to cool the PVDF.

But developing the sensor was only half the battle — umbo vibrations are so tiny that the team needed to amplify the signal without introducing too much noise. When they couldn’t find a suitable low-noise amplifier that also used very little power, they built their own.

With both prototypes in place, the researchers tested the UmboMic in human ear bones from cadavers and found that it had robust performance within the intensity and frequency range of human speech. The microphone and amplifier together also have a low noise floor, which means they could distinguish very quiet sounds from the overall noise level.

“One thing we saw that was really interesting is that the frequency response of the sensor is influenced by the anatomy of the ear we are experimenting on, because the umbo moves slightly differently in different people’s ears,” Wawrzynek says.

The researchers are preparing to launch live animal studies to further explore this finding. These experiments will also help them determine how the UmboMic responds to being implanted.

In addition, they are studying ways to encapsulate the sensor so it can remain in the body safely for up to 10 years but still be flexible enough to capture vibrations. Implants are often packaged in titanium, which would be too rigid for the UmboMic. They also plan to explore methods for mounting the UmboMic that won’t introduce vibrations.

“The results in this paper show the necessary broad-band response and low noise needed to act as an acoustic sensor. This result is surprising, because the bandwidth and noise floor are so competitive with the commercial hearing aid microphone. This performance shows the promise of the approach, which should inspire others to adopt this concept. I would expect that smaller size sensing elements and lower power electronics would be needed for next generation devices to enhance ease of implantation and battery life issues,” says Karl Grosh, professor of mechanical engineering at the University of Michigan, who was not involved with this work.

This research was funded, in part, by the National Institutes of Health, the National Science Foundation, the Cloetta Foundation in Zurich, Switzerland, and the Research Fund of the University of Basel, Switzerland.

State-of-the-art prosthetic limbs can help people with amputations achieve a natural walking gait, but they don’t give the user full neural control over the limb. Instead, they rely on robotic sensors and controllers that move the limb using predefined gait algorithms.

Using a new type of surgical intervention and neuroprosthetic interface, MIT researchers, in collaboration with colleagues from Brigham and Women’s Hospital, have shown that a natural walking gait is achievable using a prosthetic leg fully driven by the body’s own nervous system. The surgical amputation procedure reconnects muscles in the residual limb, which allows patients to receive “proprioceptive” feedback about where their prosthetic limb is in space.

In a study of seven patients who had this surgery, the MIT team found that they were able to walk faster, avoid obstacles, and climb stairs much more naturally than people with a traditional amputation.

“This is the first prosthetic study in history that shows a leg prosthesis under full neural modulation, where a biomimetic gait emerges. No one has been able to show this level of brain control that produces a natural gait, where the human’s nervous system is controlling the movement, not a robotic control algorithm,” says Hugh Herr, a professor of media arts and sciences, co-director of the K. Lisa Yang Center for Bionics at MIT, an associate member of MIT’s McGovern Institute for Brain Research, and the senior author of the new study.

Patients also experienced less pain and less muscle atrophy following this surgery, which is known as the agonist-antagonist myoneural interface (AMI). So far, about 60 patients around the world have received this type of surgery, which can also be done for people with arm amputations.

Hyungeun Song, a postdoc in MIT’s Media Lab, is the lead author of the paper, which appears today in Nature Medicine.

Sensory feedback

Most limb movement is controlled by pairs of muscles that take turns stretching and contracting. During a traditional below-the-knee amputation, the interactions of these paired muscles are disrupted. This makes it very difficult for the nervous system to sense the position of a muscle and how fast it’s contracting — sensory information that is critical for the brain to decide how to move the limb.

People with this kind of amputation may have trouble controlling their prosthetic limb because they can’t accurately sense where the limb is in space. Instead, they rely on robotic controllers built into the prosthetic limb. These limbs also include sensors that can detect and adjust to slopes and obstacles.

To try to help people achieve a natural gait under full nervous system control, Herr and his colleagues began developing the AMI surgery several years ago. Instead of severing natural agonist-antagonist muscle interactions, they connect the two ends of the muscles so that they still dynamically communicate with each other within the residual limb. This surgery can be done during a primary amputation, or the muscles can be reconnected after the initial amputation as part of a revision procedure.

“With the AMI amputation procedure, to the greatest extent possible, we attempt to connect native agonists to native antagonists in a physiological way so that after amputation, a person can move their full phantom limb with physiologic levels of proprioception and range of movement,” Herr says.

In a 2021 study, Herr’s lab found that patients who had this surgery were able to more precisely control the muscles of their amputated limb, and that those muscles produced electrical signals similar to those from their intact limb.

After those encouraging results, the researchers set out to explore whether those electrical signals could generate commands for a prosthetic limb and at the same time give the user feedback about the limb’s position in space. The person wearing the prosthetic limb could then use that proprioceptive feedback to volitionally adjust their gait as needed.

In the new Nature Medicine study, the MIT team found this sensory feedback did indeed translate into a smooth, near-natural ability to walk and navigate obstacles.

“Because of the AMI neuroprosthetic interface, we were able to boost that neural signaling, preserving as much as we could. This was able to restore a person's neural capability to continuously and directly control the full gait, across different walking speeds, stairs, slopes, even going over obstacles,” Song says.

A natural gait

For this study, the researchers compared seven people who had the AMI surgery with seven who had traditional below-the-knee amputations. All of the subjects used the same type of bionic limb: a prosthesis with a powered ankle as well as electrodes that can sense electromyography (EMG) signals from the tibialis anterior the gastrocnemius muscles. These signals are fed into a robotic controller that helps the prosthesis calculate how much to bend the ankle, how much torque to apply, or how much power to deliver.

The researchers tested the subjects in several different situations: level-ground walking across a 10-meter pathway, walking up a slope, walking down a ramp, walking up and down stairs, and walking on a level surface while avoiding obstacles.

In all of these tasks, the people with the AMI neuroprosthetic interface were able to walk faster — at about the same rate as people without amputations — and navigate around obstacles more easily. They also showed more natural movements, such as pointing the toes of the prosthesis upward while going up stairs or stepping over an obstacle, and they were better able to coordinate the movements of their prosthetic limb and their intact limb. They were also able to push off the ground with the same amount of force as someone without an amputation.

“With the AMI cohort, we saw natural biomimetic behaviors emerge,” Herr says. “The cohort that didn’t have the AMI, they were able to walk, but the prosthetic movements weren’t natural, and their movements were generally slower.”

These natural behaviors emerged even though the amount of sensory feedback provided by the AMI was less than 20 percent of what would normally be received in people without an amputation.

“One of the main findings here is that a small increase in neural feedback from your amputated limb can restore significant bionic neural controllability, to a point where you allow people to directly neurally control the speed of walking, adapt to different terrain, and avoid obstacles,” Song says.

“This work represents yet another step in us demonstrating what is possible in terms of restoring function in patients who suffer from severe limb injury. It is through collaborative efforts such as this that we are able to make transformational progress in patient care,” says Matthew Carty, a surgeon at Brigham and Women’s Hospital and associate professor at Harvard Medical School, who is also an author of the paper.

Enabling neural control by the person using the limb is a step toward Herr’s lab’s goal of “rebuilding human bodies,” rather than having people rely on ever more sophisticated robotic controllers and sensors — tools that are powerful but do not feel like part of the user’s body.

“The problem with that long-term approach is that the user would never feel embodied with their prosthesis. They would never view the prosthesis as part of their body, part of self,” Herr says. “The approach we’re taking is trying to comprehensively connect the brain of the human to the electromechanics.”

The research was funded by the MIT K. Lisa Yang Center for Bionics and the Eunice Kennedy Shriver National Institute of Child Health and Human Development.

A material with a high electron mobility is like a highway without traffic. Any electrons that flow into the material experience a commuter’s dream, breezing through without any obstacles or congestion to slow or scatter them off their path.

The higher a material’s electron mobility, the more efficient its electrical conductivity, and the less energy is lost or wasted as electrons zip through. Advanced materials that exhibit high electron mobility will be essential for more efficient and sustainable electronic devices that can do more work with less power.

Now, physicists at MIT, the Army Research Lab, and elsewhere have achieved a record-setting level of electron mobility in a thin film of ternary tetradymite — a class of mineral that is naturally found in deep hydrothermal deposits of gold and quartz.

For this study, the scientists grew pure, ultrathin films of the material, in a way that minimized defects in its crystalline structure. They found that this nearly perfect film — much thinner than a human hair — exhibits the highest electron mobility in its class.

The team was able to estimate the material’s electron mobility by detecting quantum oscillations when electric current passes through. These oscillations are a signature of the quantum mechanical behavior of electrons in a material. The researchers detected a particular rhythm of oscillations that is characteristic of high electron mobility — higher than any ternary thin films of this class to date.

“Before, what people had achieved in terms of electron mobility in these systems was like traffic on a road under construction — you’re backed up, you can’t drive, it’s dusty, and it’s a mess,” says Jagadeesh Moodera, a senior research scientist in MIT’s Department of Physics. “In this newly optimized material, it’s like driving on the Mass Pike with no traffic.”

The team’s results, which appear today in the journal Materials Today Physics, point to ternary tetradymite thin films as a promising material for future electronics, such as wearable thermoelectric devices that efficiently convert waste heat into electricity. (Tetradymites are the active materials that cause the cooling effect in commercial thermoelectric coolers.) The material could also be the basis for spintronic devices, which process information using an electron’s spin, using far less power than conventional silicon-based devices.

The study also uses quantum oscillations as a highly effective tool for measuring a material’s electronic performance.

“We are using this oscillation as a rapid test kit,” says study author Hang Chi, a former research scientist at MIT who is now at the University of Ottawa. “By studying this delicate quantum dance of electrons, scientists can start to understand and identify new materials for the next generation of technologies that will power our world.”

Chi and Moodera’s co-authors include Patrick Taylor, formerly of MIT Lincoln Laboratory, along with Owen Vail and Harry Hier of the Army Research Lab, and Brandi Wooten and Joseph Heremans of Ohio State University.

Beam down

The name “tetradymite” derives from the Greek “tetra” for “four,” and “dymite,” meaning “twin.” Both terms describe the mineral’s crystal structure, which consists of rhombohedral crystals that are “twinned” in groups of four — i.e. they have identical crystal structures that share a side.

Tetradymites comprise combinations of bismuth, antimony tellurium, sulfur, and selenium. In the 1950s, scientists found that tetradymites exhibit semiconducting properties that could be ideal for thermoelectric applications: The mineral in its bulk crystal form was able to passively convert heat into electricity.

Then, in the 1990s, the late Institute Professor Mildred Dresselhaus proposed that the mineral’s thermoelectric properties might be significantly enhanced, not in its bulk form but within its microscopic, nanometer-scale surface, where the interactions of electrons is more pronounced. (Heremans happened to work in Dresselhaus’ group at the time.)

“It became clear that when you look at this material long enough and close enough, new things will happen,” Chi says. “This material was identified as a topological insulator, where scientists could see very interesting phenomena on their surface. But to keep uncovering new things, we have to master the material growth.”

To grow thin films of pure crystal, the researchers employed molecular beam epitaxy — a method by which a beam of molecules is fired at a substrate, typically in a vacuum, and with precisely controlled temperatures. When the molecules deposit on the substrate, they condense and build up slowly, one atomic layer at a time. By controlling the timing and type of molecules deposited, scientists can grow ultrathin crystal films in exact configurations, with few if any defects.

“Normally, bismuth and tellurium can interchange their position, which creates defects in the crystal,” co-author Taylor explains. “The system we used to grow these films came down with me from MIT Lincoln Laboratory, where we use high purity materials to minimize impurities to undetectable limits. It is the perfect tool to explore this research.”

Free flow

The team grew thin films of ternary tetradymite, each about 100 nanometers thin. They then tested the film’s electronic properties by looking for Shubnikov-de Haas quantum oscillations — a phenomenon that was discovered by physicists Lev Shubnikov and Wander de Haas, who found that a material’s electrical conductivity can oscillate when exposed to a strong magnetic field at low temperatures. This effect occurs because the material’s electrons fill up specific energy levels that shift as the magnetic field changes.

Such quantum oscillations could serve as a signature of a material’s electronic structure, and the ways in which electrons behave and interact. Most notably for the MIT team, the oscillations could determine a material’s electron mobility: If oscillations exist, it must mean that the material’s electrical resistance is able to change, and by inference, electrons can be mobile, and made to easily flow.

The team looked for signs of quantum oscillations in their new films, by first exposing them to ultracold temperatures and a strong magnetic field, then running an electric current through the film and measuring the voltage along its path, as they tuned the magnetic field up and down.

“It turns out, to our great joy and excitement, that the material’s electrical resistance oscillates,” Chi says. “Immediately, that tells you that this has very high electron mobility.”

Specifically, the team estimates that the ternary tetradymite thin film exhibits an electron mobility of 10,000 cm²/V-s — the highest mobility of any ternary tetradymite film yet measured. The team suspects that the film’s record mobility has something to do with its low defects and impurities, which they were able to minimize with their precise growth strategies. The fewer a material’s defects, the fewer obstacles an electron encounters, and the more freely it can flow.

“This is showing it’s possible to go a giant step further, when properly controlling these complex systems,” Moodera says. “This tells us we’re in the right direction, and we have the right system to proceed further, to keep perfecting this material down to even much thinner films and proximity coupling for use in future spintronics and wearable thermoelectric devices.”

This research was supported in part by the Army Research Office, National Science Foundation, Office of Naval Research, Canada Research Chairs Program and Natural Sciences and Engineering Research Council of Canada.

Artificial intelligence models often play a role in medical diagnoses, especially when it comes to analyzing images such as X-rays. However, studies have found that these models don’t always perform well across all demographic groups, usually faring worse on women and people of color.

These models have also been shown to develop some surprising abilities. In 2022, MIT researchers reported that AI models can make accurate predictions about a patient’s race from their chest X-rays — something that the most skilled radiologists can’t do.

That research team has now found that the models that are most accurate at making demographic predictions also show the biggest “fairness gaps” — that is, discrepancies in their ability to accurately diagnose images of people of different races or genders. The findings suggest that these models may be using “demographic shortcuts” when making their diagnostic evaluations, which lead to incorrect results for women, Black people, and other groups, the researchers say.

“It’s well-established that high-capacity machine-learning models are good predictors of human demographics such as self-reported race or sex or age. This paper re-demonstrates that capacity, and then links that capacity to the lack of performance across different groups, which has never been done,” says Marzyeh Ghassemi, an MIT associate professor of electrical engineering and computer science, a member of MIT’s Institute for Medical Engineering and Science, and the senior author of the study.

The researchers also found that they could retrain the models in a way that improves their fairness. However, their approached to “debiasing” worked best when the models were tested on the same types of patients they were trained on, such as patients from the same hospital. When these models were applied to patients from different hospitals, the fairness gaps reappeared.

“I think the main takeaways are, first, you should thoroughly evaluate any external models on your own data because any fairness guarantees that model developers provide on their training data may not transfer to your population. Second, whenever sufficient data is available, you should train models on your own data,” says Haoran Zhang, an MIT graduate student and one of the lead authors of the new paper. MIT graduate student Yuzhe Yang is also a lead author of the paper, which appears today in Nature Medicine. Judy Gichoya, an associate professor of radiology and imaging sciences at Emory University School of Medicine, and Dina Katabi, the Thuan and Nicole Pham Professor of Electrical Engineering and Computer Science at MIT, are also authors of the paper.

Removing bias

As of May 2024, the FDA has approved 882 AI-enabled medical devices, with 671 of them designed to be used in radiology. Since 2022, when Ghassemi and her colleagues showed that these diagnostic models can accurately predict race, they and other researchers have shown that such models are also very good at predicting gender and age, even though the models are not trained on those tasks.

“Many popular machine learning models have superhuman demographic prediction capacity — radiologists cannot detect self-reported race from a chest X-ray,” Ghassemi says. “These are models that are good at predicting disease, but during training are learning to predict other things that may not be desirable.”

In this study, the researchers set out to explore why these models don’t work as well for certain groups. In particular, they wanted to see if the models were using demographic shortcuts to make predictions that ended up being less accurate for some groups. These shortcuts can arise in AI models when they use demographic attributes to determine whether a medical condition is present, instead of relying on other features of the images.

Using publicly available chest X-ray datasets from Beth Israel Deaconess Medical Center in Boston, the researchers trained models to predict whether patients had one of three different medical conditions: fluid buildup in the lungs, collapsed lung, or enlargement of the heart. Then, they tested the models on X-rays that were held out from the training data.

Overall, the models performed well, but most of them displayed “fairness gaps” — that is, discrepancies between accuracy rates for men and women, and for white and Black patients.

The models were also able to predict the gender, race, and age of the X-ray subjects. Additionally, there was a significant correlation between each model’s accuracy in making demographic predictions and the size of its fairness gap. This suggests that the models may be using demographic categorizations as a shortcut to make their disease predictions.

The researchers then tried to reduce the fairness gaps using two types of strategies. For one set of models, they trained them to optimize “subgroup robustness,” meaning that the models are rewarded for having better performance on the subgroup for which they have the worst performance, and penalized if their error rate for one group is higher than the others.

In another set of models, the researchers forced them to remove any demographic information from the images, using “group adversarial” approaches. Both strategies worked fairly well, the researchers found.

“For in-distribution data, you can use existing state-of-the-art methods to reduce fairness gaps without making significant trade-offs in overall performance,” Ghassemi says. “Subgroup robustness methods force models to be sensitive to mispredicting a specific group, and group adversarial methods try to remove group information completely.”

Not always fairer

However, those approaches only worked when the models were tested on data from the same types of patients that they were trained on — for example, only patients from the Beth Israel Deaconess Medical Center dataset.

When the researchers tested the models that had been “debiased” using the BIDMC data to analyze patients from five other hospital datasets, they found that the models’ overall accuracy remained high, but some of them exhibited large fairness gaps.

“If you debias the model in one set of patients, that fairness does not necessarily hold as you move to a new set of patients from a different hospital in a different location,” Zhang says.

This is worrisome because in many cases, hospitals use models that have been developed on data from other hospitals, especially in cases where an off-the-shelf model is purchased, the researchers say.

“We found that even state-of-the-art models which are optimally performant in data similar to their training sets are not optimal — that is, they do not make the best trade-off between overall and subgroup performance — in novel settings,” Ghassemi says. “Unfortunately, this is actually how a model is likely to be deployed. Most models are trained and validated with data from one hospital, or one source, and then deployed widely.”

The researchers found that the models that were debiased using group adversarial approaches showed slightly more fairness when tested on new patient groups than those debiased with subgroup robustness methods. They now plan to try to develop and test additional methods to see if they can create models that do a better job of making fair predictions on new datasets.

The findings suggest that hospitals that use these types of AI models should evaluate them on their own patient population before beginning to use them, to make sure they aren’t giving inaccurate results for certain groups.

The research was funded by a Google Research Scholar Award, the Robert Wood Johnson Foundation Harold Amos Medical Faculty Development Program, RSNA Health Disparities, the Lacuna Fund, the Gordon and Betty Moore Foundation, the National Institute of Biomedical Imaging and Bioengineering, and the National Heart, Lung, and Blood Institute.

Researchers from MIT and the University of Michigan have discovered a new way to drive chemical reactions that could generate a wide variety of compounds with desirable pharmaceutical properties.

These compounds, known as azetidines, are characterized by four-membered rings that include nitrogen. Azetidines have traditionally been much more difficult to synthesize than five-membered nitrogen-containing rings, which are found in many FDA-approved drugs.

The reaction that the researchers used to create azetidines is driven by a photocatalyst that excites the molecules from their ground energy state. Using computational models that they developed, the researchers were able to predict compounds that can react with each other to form azetidines using this kind of catalysis.

“Going forward, rather than using a trial-and-error process, people can prescreen compounds and know beforehand which substrates will work and which ones won't,” says Heather Kulik, an associate professor of chemistry and chemical engineering at MIT.

Kulik and Corinna Schindler, a professor of chemistry at the University of Michigan, are the senior authors of the study, which appears today in Science. Emily Wearing, recently a graduate student at the University of Michigan, is the lead author of the paper. Other authors include University of Michigan postdoc Yu-Cheng Yeh, MIT graduate student Gianmarco Terrones, University of Michigan graduate student Seren Parikh, and MIT postdoc Ilia Kevlishvili.

Light-driven synthesis

Many naturally occurring molecules, including vitamins, nucleic acids, enzymes and hormones, contain five-membered nitrogen-containing rings, also known as nitrogen heterocycles. These rings are also found in more than half of all FDA-approved small-molecule drugs, including many antibiotics and cancer drugs.

Four-membered nitrogen heterocycles, which are rarely found in nature, also hold potential as drug compounds. However, only a handful of existing drugs, including penicillin, contain four-membered heterocycles, in part because these four-membered rings are much more difficult to synthesize than five-membered heterocycles.

In recent years, Schindler’s lab has been working on synthesizing azetidines using light to drive a reaction that combines two precursors, an alkene and an oxime. These reactions require a photocatalyst, which absorbs light and passes the energy to the reactants, making it possible for them to react with each other.

“The catalyst can transfer that energy to another molecule, which moves the molecules into excited states and makes them more reactive. This is a tool that people are starting to use to make it possible to make certain reactions occur that wouldn't normally occur,” Kulik says.

Schindler’s lab found that while this reaction sometimes worked well, other times it did not, depending on which reactants were used. They enlisted Kulik, an expert in developing computational approaches to modeling chemical reactions, to help them figure out how to predict when these reactions will occur.

The two labs hypothesized that whether a particular alkene and oxime will react together in a photocatalyzed reaction depends on a property known as the frontier orbital energy match. Electrons that surround the nucleus of an atom exist in orbitals, and quantum mechanics can be used to predict the shape and energies of these orbitals. For chemical reactions, the most important electrons are those in the outermost, highest energy (“frontier”) orbitals, which are available to react with other molecules.

Kulik and her students used density functional theory, which uses the Schrödinger equation to predict where electrons could be and how much energy they have, to calculate the orbital energy of these outermost electrons.

These energy levels are also affected by other groups of atoms attached to the molecule, which can change the properties of the electrons in the outermost orbitals.

Once those energy levels are calculated, the researchers can identify reactants that have similar energy levels when the photocatalyst boosts them into an excited state. When the excited states of an alkene and an oxime are closely matched, less energy is required to boost the reaction to its transition state — the point at which the reaction has enough energy to go forward to form products.

Accurate predictions

After calculating the frontier orbital energies for 16 different alkenes and nine oximes, the researchers used their computational model to predict whether 18 different alkene-oxime pairs would react together to form an azetidine. With the calculations in hand, these predictions can be made in a matter of seconds.

The researchers also modeled a factor that influences the overall yield of the reaction: a measure of how available the carbon atoms in the oxime are to participate in chemical reactions.

The model’s predictions suggested that some of these 18 reactions won’t occur or won’t give a high enough yield. However, the study also showed that a significant number of reactions are correctly predicted to work.

“Based on our model, there's a much wider range of substrates for this azetidine synthesis than people thought before. People didn't really think that all of this was accessible,” Kulik says.

Of the 27 combinations that they studied computationally, the researchers tested 18 reactions experimentally, and they found that most of their predictions were accurate. Among the compounds they synthesized were derivatives of two drug compounds that are currently FDA-approved: amoxapine, an antidepressant, and indomethacin, a pain reliever used to treat arthritis.

This computational approach could help pharmaceutical companies predict molecules that will react together to form potentially useful compounds, before spending a lot of money to develop a synthesis that might not work, Kulik says. She and Schindler are continuing to work together on other kinds of novel syntheses, including the formation of compounds with three-membered rings.

“Using photocatalysts to excite substrates is a very active and hot area of development, because people have exhausted what you can do on the ground state or with radical chemistry,” Kulik says. “I think this approach is going to have a lot more applications to make molecules that are normally thought of as really challenging to make.”