Well into the late 19th century, the U.S. retail sector was overwhelmingly local, consisting of small, independent merchants throughout the country. That started changing after Sears and Roebuck’s famous catalog became popular, allowing the firm to grow, while a rival, Montgomery Ward, also expanded. By the 1930s, the U.S. had 130,000 chain stores, topped by Atlantic and Pacific supermarkets (the A&P), with over 15,000 stores.

A century onward, the U.S. retail landscape is dominated by retail giants. Today, 90 percent of Americans live within 10 miles of a Walmart, while five of the country’s 10 biggest employers — Walmart, Amazon, Home Depot, Kroger, and Target— are retailers. Two others in the top 10, UPS and FedEx, are a major part of the retail economy.

The ubiquity of these big retailers, and the sheer extent of the U.S. shopping economy as a whole, is unusual compared to the country’s European counterparts. Domestic consumption plays an outsized role in driving growth in the United States, and credit plays a much larger role in supporting that consumption than in Europe. The U.S. has five times as much retail space per capita as Japan and the U.K., and 10 times as much as Germany. Unlike in Europe, shopping hours are largely unregulated.

How did this happen? To be sure, Walmart, Amazon, Target, and other massive chains have plenty of business acumen. But the full story involves a century or more of political tectonics and legal debates, which helped shape the size of U.S. retailing and the prominence of its large discount chains. 

“The markets that we take as given, that we think of as the natural outcome of supply and demand, are heavily shaped by policy and by politics,” says MIT political scientist Kathleen Thelen.

Thelen examines the subject in a new book, “Attention, Shoppers! American Retail Capitalism and the Origins of the Amazon Economy,” published today by Princeton University Press. In it, she examines the growth of the particular model of supersized, low-cost, low-wage retailing that now features so prominently in the U.S. economy.

Prioritizing prices

While a great deal has been written about specific American companies, Thelen’s book has some distinctive features. One is a comparison to the economies of Europe, where she has focused much of her scholarship. Another is her historical lens, extending back to the start of chain retailing.

“It seems like every time I set out to explain something in the present, I’m thrown back to the 19th century,” Thelen says.

For instance, as both Sears and Montgomery Ward grew, producers and consumers were still experimenting with alternate commercial arrangements, like cooperatives, which pooled suppliers together, but they ultimately ran into economic and legal headwinds. Especially, at the time, legal headwinds.

“Antitrust laws in the United States were very forbearing toward big multidivisional corporations and very punitive toward alternative types of arrangements like cooperatives, so big retailers got a real boost in that period,” Thelen says. Separately, the U.S. Postal Service was also crucial, since big mail order houses like Sears relied on not just on its delivery services but also its money order system, to sell goods to the company’s many customers who lacked bank accounts.

Smaller retailers fought large chains during the Depression, especially in the South and the West, which forms another phase of the story. But low-cost discounters worked around some laws through regulatory arbitrage, finding friendlier regulations in some states — and sometimes though outright rule-breaking. Ultimately, larger retailers have thrived again in the last half century, especially as antitrust law increasingly prioritized consumer prices as its leading measuring stick.

Most antitrust theorizing since the 1960s “valorizes consumer welfare, which is basically defined as price, so anything that delivers the lowest price to consumers is A-OK,” Thelen says. “We’re in this world where the large, low-cost retailers are delivering consumer welfare in the way the courts are defining it.”

That emphasis on prices, she notes, then spills over into other areas of the economy, especially wages and labor relations.

“If you prioritize prices, one of the main ways to reduce prices is to reduce labor costs,” Thelen says. “It’s no coincidence that low-cost discounters are often low-wage employers. Indeed, they often squeeze their vendors to deliver goods at ever-lower prices, and by extension they’re pressing down on wages in their supplier networks as well.”

As Thelen’s book explains, legal views supporting large chains were also common during the first U.S. wave of chain-retail growth. She writes, “large, low-cost retailers have almost always enjoyed a privileged position in the American antitrust regime.”

In the “deep equilibrium”

“Attention, Shoppers!” makes clear that this tendency toward lower prices, lower employee pay, and high consumer convenience is particularly pronounced in the U.S., where 22.6 percent of employees count as low-wage workers (making two-thirds or less of the country’s median wage). In the other countries that belong to the Organization for Economic Cooperation and Development, 13.9 percent of workers fit that description. About three-quarters of U.S. retail workers are in the low-wage category.

In other OECD countries, on aggregate, manufacturers and producers make up bigger chunks of the economy and, correspondingly, often have legal frameworks more friendly to manufacturers and to labor. But in the U.S., large retailers have gained more leverage, if anything, in the last half-century, Thelen notes.

“You might think mass retailers and manufacturers would have a symbiotic relationship, but historically there has been great tension between them, especially on price,” Thelen says. “In the postwar period, the balance of power became tilted toward retailers, and away from manufacturers and labor. Retailers also had consumers on their side, and had more power over data to dictate the terms on which their vendors would supply goods to them.”

Currently, as Thelen writes in the book, the U.S. is in a “deep equilibrium” on this front, in that many low-wage workers now rely on these low-cost retailers to make ends meet — and because Americans as a whole now find it normal to have their purchases delivered at lightning speed. Things might be different, Thelen suggests, if there are changes to U.S. antitrust enforcement, or, especially, major reforms to labor law, such as allowing workers to organize for higher wages across companies, not just at individual stores. Short of that, the equilibrium is likely to hold.

“Attention, Shoppers!” has received praise from other scholars. Louis Hyman, a historian at Johns Hopkins University, has called it a “pathbreaking study that provides insight into not only the past but also the future of online retail.”

For her part, Thelen hopes readers will learn more about an economic landscape we might take for granted, even while we shop at big chains, around us and online.

“The triumph of these types of retailers was not inevitable,” Thelen says. “It was a function of politics and political choice.”

Think of your most prized belongings. In an increasingly virtual world, wouldn’t it be great to save a copy of that precious item and all the memories it holds?

In mixed-reality settings, you can create a digital twin of a physical item, such as an old doll. But it’s hard to replicate interactive elements, like the way it moves or the sounds it makes — the sorts of unique interactive features that made the toy distinct in the first place.

Researchers from MIT’s Computer Science and Artificial Intelligence Laboratory (CSAIL) sought to change that, and they have a potential solution. Their “InteRecon” program enables users to recapture real-world objects in a mobile app, and then animate them in mixed-reality environments. 

This prototype could recreate the interaction functions in the physical world, such as the head motions of your favorite bobblehead, or playing a classic video on a digital version of your vintage TV. It creates more lifelike and personal digital surroundings while preserving a memory.

InteRecon’s ability to reconstruct the interactive experience of different items could make it a useful tool for teachers explaining important concepts, like demonstrating how gravity pulls an object down. It could also add a new visual component to museum exhibits, such as animating a painting or bringing a historical mannequin to life (without the scares of characters from “Night at the Museum”). Eventually, InteRecon may be able to teach a doctor’s apprentice organ surgery or a cosmetic procedure by visualizing each motion needed to complete the task.

The exciting potential of InteRecon comes from its ability to add motions or interactive functions to many different objects, according to CSAIL visiting researcher Zisu Li, lead author of a paper introducing the tool.

“While taking a picture or video is a great way to preserve a memory, those digital copies are static,” says Li, who is also a PhD student at the Hong Kong University of Science and Technology. “We found that users wanted to reconstruct personal items while preserving their interactivity to enrich their memories. With the power of mixed reality, InteRecon can make these memories live longer in virtual settings as interactive digital items.”

Li and her colleagues will present InteRecon at the 2025 ACM CHI conference on Human Factors in Computing Systems.

Making a virtual world more realistic

To make digital interactivity possible, the team first developed an iPhone app. Using your camera, you scan the item all the way around three times to ensure it’s fully captured. The 3D model can then be imported into the InteRecon mixed reality interface, where you can mark (“segment”) individual areas to select which parts of the model will be interactive (like a doll’s arms, head, torso, and legs). Alternatively, you can use the function provided by InteRecon for automatic segmentation.

The InteRecon interface can be accessed via the mixed reality headset (such as Hololens 2 and Quest). It allows you to choose a programmable motion for the part of the item you want to animate after your model is segmented.

Movement options are presented as motion demonstrations, allowing you to play around with them before deciding on one — say, a flopping motion that emulates how a bunny doll’s ears move. You can even pinch a specific part and explore different ways to animate it, like sliding, dangling, and pendulum-like turns.

Your old iPod, digitized

The team showed that InteRecon can also recapture the interface of physical electronic devices, like a vintage TV. After making a digital copy of the item, you can customize the 3D model with different interfaces.

Users can play with example widgets from different interfaces before choosing a motion: a screen (either a TV display or camera’s viewfinder), a rotating knob (for, say, adjusting the volume), an “on/off”-style button, and a slider (for changing settings on something like a DJ booth).

Li and colleagues presented an application that recreates the interactivity of a vintage TV by incorporating virtual widgets such as an “on/off” button, a screen, and a channel switch on a TV model, along with embedding old videos into it. This makes the TV model come to life. You could also upload MP3 files and add a “play button” to a 3D model of an iPod to listen to your favorite songs in mixed reality.

The researchers believe InteRecon opens up intriguing new avenues in designing lifelike virtual environments. A user study confirmed that people from different fields share this enthusiasm, viewing it as easy to learn and diverse in its ability to express the richness of users’ memories.

“One thing I really appreciate is that the items that users remember are imperfect,” says Faraz Faruqi SM ’22, another author on the paper who is also a CSAIL affiliate and MIT PhD student in electrical engineering and computer science. “InteRecon brings those imperfections into mixed reality, accurately recreating what made a personal item like a teddy bear missing a few buttons so special.”

In a related study, users imagined how this technology could be applied to professional scenarios, from teaching medical students how to perform surgeries to helping travelers and researchers log their trips, and even assisting fashion designers in experimenting with materials.

Before InteRecon is used in more advanced settings, though, the team would like to upgrade their physical simulation engine to something more precise. This would enable applications such as helping a doctor’s apprentice to learn the pinpoint accuracy needed to do certain surgical maneuvers.

Li and Faruqi may also incorporate large language models and generative models that can recreate lost personal items into 3D models via language descriptions, as well as explain the interface’s features.

As for the researchers’ next steps, Li is working toward a more automatic and powerful pipeline that can make interactivity-preserved digital twins of larger physical environments in mixed reality for end users, such as a virtual office space. Faruqi is looking to build an approach that can physically recreate lost items via 3D printers.

“InteRecon represents an exciting new frontier in the field of mixed reality, going beyond mere visual replication to capture the unique interactivity of physical objects,” says Hanwang Zhang, an associate professor at Nanyang Technological University's College of Computing and Data Science, who wasn’t involved in the research. “This technology has the potential to revolutionize education, health care, and cultural exhibitions by bringing a new level of immersion and personal connection to virtual environments.”

Li and Faruqi wrote the paper with the Hong Kong University of Science and Technology (HKUST) master’s student Jiawei Li, PhD student Shumeng Zhang, Associate Professor Xiaojuan Ma, and assistant professors Mingming Fan and Chen Liang from HKUST; ETH Zurich PhD student Zeyu Xiong; and 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. Their work was supported by the APEX Lab of The Hong Kong University of Science and Technology (Guangzhou) in collaboration with the HCI Engineering Group.

Watching and listening to a pianist’s performance is an immersive and enjoyable experience. The pianist and the instrument, with a blend of skill, training, and presence, create a series of memorable moments for themselves and the audience. But is there a way to improve the performance and our understanding of how the performer and their instrument work together to create this magic, while also minimizing performance-related injuries?

Mi-Eun Kim, director of keyboard studies in MIT’s Music and Theater Arts Section, and Praneeth Namburi PhD ’16, a research scientist in MIT’s Institute for Medical Engineering and Science, are investigating how the body works when pianists play. Their joint project, The Biomechanics of Assimilating a New Piano Skill, aims to develop mechanistic insights that could transform how we understand and teach piano technique, reduce performance-related injuries, and bridge the gap between artistic expression and biomechanical efficiency. 

Their project is among those recently selected for a SHASS+ Connectivity Fund grant through the MIT Human Insight Collaborative.

“The project emerged from a convergence of interests and personal experiences,” Namburi says. “Mi-Eun witnessed widespread injuries among fellow pianists and saw how these injuries could derail careers.”

Kim is a renowned pianist who has performed on stages throughout the United States, in Europe, and in Asia. She earned the Liszt-Garrison Competition’s Liszt Award and the Corpus Christi solo prize, among other honors. She teaches piano and chamber music through MIT Music’s Emerson/Harris Program and chamber music through MIT’s Chamber Music Society. She earned advanced degrees from the University of Michigan and holds a bachelor of arts degree in history from Columbia University.

Namburi’s work focuses on the biomechanics of efficient, expressive, and coordinated movement. He draws inspiration from artists and athletes in specialized movement disciplines, such as dancing and fencing, to investigate skilled movement. He earned a PhD in experimental neuroscience from MIT and a bachelor of engineering degree in electrical and electronic engineering from Singapore’s Nanyang Technological University. 

Pursuing the project

Kim and Namburi arrived at their project by taking different roads into the arts. While Kim was completing her studies at the University of Michigan, Namburi was taking dance lessons as a hobby in Boston. He learned that both expressive and sustainable movements might share a common denominator. “A key insight was that elastic tissues play a crucial role in coordinated, expressive, and sustainable movements in dance — a principle that could extend beyond dancing,” he notes.

“We recognized that studying elastic tissues could shed light on reducing injury risk, as well as understanding musical expression and embodiment in the context of piano playing,” Kim says.

Kim and Namburi began collaborating on what would become their project in October 2023, though the groundwork was in place months before. “A visiting student working with me on a research project studying pianists in the MIT.nano Immersion Lab reached out to Mi-Eun in summer 2023,” Namburi recalls. A shared Instagram video showing their setup with motion capture sensors and a pianist playing Chopin on a digital keyboard sparked Kim’s interest. The Immersion Lab is an open-access, shared facility for MIT and beyond dedicated to visualizing, understanding, and interacting with large, multidimensional data.

“I couldn't make sense of all the sensors, but immediately noticed they were using a digital keyboard,” she says.

Kim wanted to elevate these studies’ quality by pairing the musicians with the proper equipment and instrument. While the digital pianos they’d previously used are portable and provide musical instrument digital interface (MIDI) data, they don’t offer the same experience as a real piano. “Pianists dream of playing on an ideal instrument — a 9-foot concert grand with perfectly regulated 24-inch keys that responds to every musical intention without resistance,” Kim says.

The researchers brought a Spirio grand piano to the Immersion Lab and observed that the instrument could both capture pianists’ hammerstrike velocities and reproduce them to play back the performance. Monitoring Kim’s performance on the concert grand piano, for example, both noted marked differences in her playing style.

“Despite all the sensors, lighting, and observers, playing felt so natural that I forgot I was in a lab,” she says. “I could focus purely on the music, without worrying about adapting to a smaller keyboard or digital sound.”

This setup allowed them to observe pianists’ natural movements, which was exactly what Kim wanted to study.

During Independent Activities Period 2025, Kim and Namburi hosted a new course, Biomechanics of Piano Playing, in the Immersion Lab. Students and faculty from MIT, Harvard University, the University of Michigan, the University of Toronto, and the University of Hartford took part. Participants learned how to use motion capture, accelerometers, and ultrasound imaging to visualize signals from the body during piano playing.

Observations and outcomes

If the efficiency and perceived fluency of an expert pianist’s movements comes from harnessing the body’s inherent elastic mechanisms, Kim and Namburi believe, it’s possible to redesign how piano playing is taught. Each wants to reduce occurrences of playing-related injuries and improve how musicians learn their craft.

“I want us to bridge the gap between artistic expression and biomechanical efficiency,” Namburi says.

Through their exploratory sessions at the Immersion Lab, Kim and Namburi found common ground, gathering information about their observations of and experiences in piano and dance through sensor technology, including ultrasound.

Beyond these, Kim saw potential for transforming piano pedagogy. “Traditional teaching relies heavily on subjective descriptions and metaphors passed down through generations,” she says. “While valuable, these approaches could be enhanced with objective, scientific understanding of the physical mechanisms behind skilled piano performance — evidence-driven piano pedagogy, if you will.”

Immune molecules called cytokines play important roles in the body’s defense against infection, helping to control inflammation and coordinating the responses of other immune cells. A growing body of evidence suggests that some of these molecules also influence the brain, leading to behavioral changes during illness.

Two new studies from MIT and Harvard Medical School, focused on a cytokine called IL-17, now add to that evidence. The researchers found that IL-17 acts on two distinct brain regions — the amygdala and the somatosensory cortex — to exert two divergent effects. In the amygdala, IL-17 can elicit feelings of anxiety, while in the cortex it promotes sociable behavior.

These findings suggest that the immune and nervous systems are tightly interconnected, says Gloria Choi, an associate professor of brain and cognitive sciences, a member of MIT’s Picower Institute for Learning and Memory, and one of the senior authors of the studies.

“If you’re sick, there’s so many more things that are happening to your internal states, your mood, and your behavioral states, and that’s not simply you being fatigued physically. It has something to do with the brain,” she says.

Jun Huh, an associate professor of immunology at Harvard Medical School, is also a senior author of both studies, which appear today in Cell. One of the papers was led by Picower Institute Research Scientist Byeongjun Lee and former Picower Institute research scientist Jeong-Tae Kwon, and the other was led by Harvard Medical School postdoc Yunjin Lee and Picower Institute postdoc Tomoe Ishikawa.

Behavioral effects

Choi and Huh became interested in IL-17 several years ago, when they found it was involved in a phenomenon known as the fever effect. Large-scale studies of autistic children have found that for many of them, their behavioral symptoms temporarily diminish when they have a fever.

In a 2019 study in mice, Choi and Huh showed that in some cases of infection, IL-17 is released and suppresses a small region of the brain’s cortex known as S1DZ. Overactivation of neurons in this region can lead to autism-like behavioral symptoms in mice, including repetitive behaviors and reduced sociability.

“This molecule became a link that connects immune system activation, manifested as a fever, to changes in brain function and changes in the animals’ behavior,” Choi says.

IL-17 comes in six different forms, and there are five different receptors that can bind to it. In their two new papers, the researchers set out to map which of these receptors are expressed in different parts of the brain. This mapping revealed that a pair of receptors known as IL-17RA and IL-17RB is found in the cortex, including in the S1DZ region that the researchers had previously identified. The receptors are located in a population of neurons that receive proprioceptive input and are involved in controlling behavior.

When a type of IL-17 known as IL-17E binds to these receptors, the neurons become less excitable, which leads to the behavioral effects seen in the 2019 study.

“IL-17E, which we’ve shown to be necessary for behavioral mitigation, actually does act almost exactly like a neuromodulator in that it will immediately reduce these neurons’ excitability,” Choi says. “So, there is an immune molecule that’s acting as a neuromodulator in the brain, and its main function is to regulate excitability of neurons.”

Choi hypothesizes that IL-17 may have originally evolved as a neuromodulator, and later on was appropriated by the immune system to play a role in promoting inflammation. That idea is consistent with previous work showing that in the worm C. elegans, IL-17 has no role in the immune system but instead acts on neurons. Among its effects in worms, IL-17 promotes aggregation, a form of social behavior. Additionally, in mammals, IL-17E is actually made by neurons in the cortex, including S1DZ.

“There’s a possibility that a couple of forms of IL-17 perhaps evolved first and foremost to act as a neuromodulator in the brain, and maybe later were hijacked by the immune system also to act as immune modulators,” Choi says.

Provoking anxiety

In the other Cell paper, the researchers explored another brain location where they found IL-17 receptors — the amygdala. This almond-shaped structure plays an important role in processing emotions, including fear and anxiety.

That study revealed that in a region known as the basolateral amygdala (BLA), the IL-17RA and IL-17RE receptors, which work as a pair, are expressed in a discrete population of neurons. When these receptors bind to IL-17A and IL-17C, the neurons become more excitable, leading to an increase in anxiety.

The researchers also found that, counterintuitively, if animals are treated with antibodies that block IL-17 receptors, it actually increases the amount of IL-17C circulating in the body. This finding may help to explain unexpected outcomes observed in a clinical trial of a drug targeting the IL-17-RA receptor for psoriasis treatment, particularly regarding its potential adverse effects on mental health.

“We hypothesize that there’s a possibility that the IL-17 ligand that is upregulated in this patient cohort might act on the brain to induce suicide ideation, while in animals there is an anxiogenic phenotype,” Choi says.

During infections, this anxiety may be a beneficial response, keeping the sick individual away from others to whom the infection could spread, Choi hypothesizes.

“Other than its main function of fighting pathogens, one of the ways that the immune system works is to control the host behavior, to protect the host itself and also protect the community the host belongs to,” she says. “One of the ways the immune system is doing that is to use cytokines, secreted factors, to go to the brain as communication tools.”

The researchers found that the same BLA neurons that have receptors for IL-17 also have receptors for IL-10, a cytokine that suppresses inflammation. This molecule counteracts the excitability generated by IL-17, giving the body a way to shut off anxiety once it’s no longer useful.

Distinctive behaviors

Together, the two studies suggest that the immune system, and even a single family of cytokines, can exert a variety of effects in the brain.

“We have now different combinations of IL-17 receptors being expressed in different populations of neurons, in two different brain regions, that regulate very distinct behaviors. One is actually somewhat positive and enhances social behaviors, and another is somewhat negative and induces anxiogenic phenotypes,” Choi says.

Her lab is now working on additional mapping of IL-17 receptor locations, as well as the IL-17 molecules that bind to them, focusing on the S1DZ region. Eventually, a better understanding of these neuro-immune interactions may help researchers develop new treatments for neurological conditions such as autism or depression.

“The fact that these molecules are made by the immune system gives us a novel approach to influence brain function as means of therapeutics,” Choi says. “Instead of thinking about directly going for the brain, can we think about doing something to the immune system?”

The research was funded, in part, by Jeongho Kim and the Brain Impact Foundation Neuro-Immune Fund, the Simons Foundation Autism Research Initiative, the Simons Center for the Social Brain, the Marcus Foundation, the N of One: Autism Research Foundation, the Burroughs Wellcome Fund, the Picower Institute Innovation Fund, the MIT John W. Jarve Seed Fund for Science Innovation, Young Soo Perry and Karen Ha, and the National Institutes of Health.

When the International Maritime Organization enacted a mandatory cap on the sulfur content of marine fuels in 2020, with an eye toward reducing harmful environmental and health impacts, it left shipping companies with three main options.

They could burn low-sulfur fossil fuels, like marine gas oil, or install cleaning systems to remove sulfur from the exhaust gas produced by burning heavy fuel oil. Biofuels with lower sulfur content offer another alternative, though their limited availability makes them a less feasible option.

While installing exhaust gas cleaning systems, known as scrubbers, is the most feasible and cost-effective option, there has been a great deal of uncertainty among firms, policymakers, and scientists as to how “green” these scrubbers are.

Through a novel lifecycle assessment, researchers from MIT, Georgia Tech, and elsewhere have now found that burning heavy fuel oil with scrubbers in the open ocean can match or surpass using low-sulfur fuels, when a wide variety of environmental factors is considered.

The scientists combined data on the production and operation of scrubbers and fuels with emissions measurements taken onboard an oceangoing cargo ship.

They found that, when the entire supply chain is considered, burning heavy fuel oil with scrubbers was the least harmful option in terms of nearly all 10 environmental impact factors they studied, such as greenhouse gas emissions, terrestrial acidification, and ozone formation.

“In our collaboration with Oldendorff Carriers to broadly explore reducing the environmental impact of shipping, this study of scrubbers turned out to be an unexpectedly deep and important transitional issue,” says Neil Gershenfeld, an MIT professor, director of the Center for Bits and Atoms (CBA), and senior author of the study.

“Claims about environmental hazards and policies to mitigate them should be backed by science. You need to see the data, be objective, and design studies that take into account the full picture to be able to compare different options from an apples-to-apples perspective,” adds lead author Patricia Stathatou, an assistant professor at Georgia Tech, who began this study as a postdoc in the CBA.

Stathatou is joined on the paper by Michael Triantafyllou, the Henry L. and Grace Doherty Professor in Ocean Science and Engineering in the Department of Mechanical Engineering and others at the National Technical University of Athens in Greece, Naias Laboratories, and the maritime shipping firm Oldendorff Carriers. The research appears today in Environmental Science and Technology.

Slashing sulfur emissions

Heavy fuel oil, traditionally burned by bulk carriers that make up about 30 percent of the global maritime fleet, usually has a sulfur content around 2 to 3 percent. This is far higher than the International Maritime Organization’s 2020 cap of 0.5 percent in most areas of the ocean and 0.1 percent in areas near population centers or environmentally sensitive regions.

Sulfur oxide emissions contribute to air pollution and acid rain, and can damage the human respiratory system.

In 2018, fewer than 1,000 vessels employed scrubbers. After the cap went into place, higher prices of low-sulfur fossil fuels and limited availability of alternative fuels led many firms to install scrubbers so they could keep burning heavy fuel oil.

Today, more than 5,800 vessels utilize scrubbers, the majority of which are wet, open-loop scrubbers.

“Scrubbers are a very mature technology. They have traditionally been used for decades in land-based applications like power plants to remove pollutants,” Stathatou says.

A wet, open-loop marine scrubber is a huge, metal, vertical tank installed in a ship’s exhaust stack, above the engines. Inside, seawater drawn from the ocean is sprayed through a series of nozzles downward to wash the hot exhaust gases as they exit the engines.

The seawater interacts with sulfur dioxide in the exhaust, converting it to sulfates — water-soluble, environmentally benign compounds that naturally occur in seawater. The washwater is released back into the ocean, while the cleaned exhaust escapes to the atmosphere with little to no sulfur dioxide emissions.

But the acidic washwater can contain other combustion byproducts like heavy metals, so scientists wondered if scrubbers were comparable, from a holistic environmental point of view, to burning low-sulfur fuels.

Several studies explored toxicity of washwater and fuel system pollution, but none painted a full picture.

The researchers set out to fill that scientific gap.

A “well-to-wake” analysis

The team conducted a lifecycle assessment using a global environmental database on production and transport of fossil fuels, such as heavy fuel oil, marine gas oil, and very-low sulfur fuel oil. Considering the entire lifecycle of each fuel is key, since producing low-sulfur fuel requires extra processing steps in the refinery, causing additional emissions of greenhouse gases and particulate matter.

“If we just look at everything that happens before the fuel is bunkered onboard the vessel, heavy fuel oil is significantly more low-impact, environmentally, than low-sulfur fuels,” she says.

The researchers also collaborated with a scrubber manufacturer to obtain detailed information on all materials, production processes, and transportation steps involved in marine scrubber fabrication and installation.

“If you consider that the scrubber has a lifetime of about 20 years, the environmental impacts of producing the scrubber over its lifetime are negligible compared to producing heavy fuel oil,” she adds.

For the final piece, Stathatou spent a week onboard a bulk carrier vessel in China to measure emissions and gather seawater and washwater samples. The ship burned heavy fuel oil with a scrubber and low-sulfur fuels under similar ocean conditions and engine settings.

Collecting these onboard data was the most challenging part of the study.

“All the safety gear, combined with the heat and the noise from the engines on a moving ship, was very overwhelming,” she says.

Their results showed that scrubbers reduce sulfur dioxide emissions by 97 percent, putting heavy fuel oil on par with low-sulfur fuels according to that measure. The researchers saw similar trends for emissions of other pollutants like carbon monoxide and nitrous oxide.

In addition, they tested washwater samples for more than 60 chemical parameters, including nitrogen, phosphorus, polycyclic aromatic hydrocarbons, and 23 metals.

The concentrations of chemicals regulated by the IMO were far below the organization’s requirements. For unregulated chemicals, the researchers compared the concentrations to the strictest limits for industrial effluents from the U.S. Environmental Protection Agency and European Union.

Most chemical concentrations were at least an order of magnitude below these requirements.

In addition, since washwater is diluted thousands of times as it is dispersed by a moving vessel, the concentrations of such chemicals would be even lower in the open ocean.

These findings suggest that the use of scrubbers with heavy fuel oil can be considered as equal to or more environmentally friendly than low-sulfur fuels across many of the impact categories the researchers studied.

“This study demonstrates the scientific complexity of the waste stream of scrubbers. Having finally conducted a multiyear, comprehensive, and peer-reviewed study, commonly held fears and assumptions are now put to rest,” says Scott Bergeron, managing director at Oldendorff Carriers and co-author of the study.

“This first-of-its-kind study on a well-to-wake basis provides very valuable input to ongoing discussion at the IMO,” adds Thomas Klenum, executive vice president of innovation and regulatory affairs at the Liberian Registry, emphasizing the need “for regulatory decisions to be made based on scientific studies providing factual data and conclusions.”

Ultimately, this study shows the importance of incorporating lifecycle assessments into future environmental impact reduction policies, Stathatou says.

“There is all this discussion about switching to alternative fuels in the future, but how green are these fuels? We must do our due diligence to compare them equally with existing solutions to see the costs and benefits,” she adds.

This study was supported, in part, by Oldendorff Carriers.

Due to the inherent ambiguity in medical images like X-rays, radiologists often use words like “may” or “likely” when describing the presence of a certain pathology, such as pneumonia.

But do the words radiologists use to express their confidence level accurately reflect how often a particular pathology occurs in patients? A new study shows that when radiologists express confidence about a certain pathology using a phrase like “very likely,” they tend to be overconfident, and vice-versa when they express less confidence using a word like “possibly.”

Using clinical data, a multidisciplinary team of MIT researchers in collaboration with researchers and clinicians at hospitals affiliated with Harvard Medical School created a framework to quantify how reliable radiologists are when they express certainty using natural language terms.

They used this approach to provide clear suggestions that help radiologists choose certainty phrases that would improve the reliability of their clinical reporting. They also showed that the same technique can effectively measure and improve the calibration of large language models by better aligning the words models use to express confidence with the accuracy of their predictions.

By helping radiologists more accurately describe the likelihood of certain pathologies in medical images, this new framework could improve the reliability of critical clinical information.

“The words radiologists use are important. They affect how doctors intervene, in terms of their decision making for the patient. If these practitioners can be more reliable in their reporting, patients will be the ultimate beneficiaries,” says Peiqi Wang, an MIT graduate student and lead author of a paper on this research.

He is joined on the paper by senior author Polina Golland, a Sunlin and Priscilla Chou Professor of Electrical Engineering and Computer Science (EECS), a principal investigator in the MIT Computer Science and Artificial Intelligence Laboratory (CSAIL), and the leader of the Medical Vision Group; as well as Barbara D. Lam, a clinical fellow at the Beth Israel Deaconess Medical Center; Yingcheng Liu, at MIT graduate student; Ameneh Asgari-Targhi, a research fellow at Massachusetts General Brigham (MGB); Rameswar Panda, a research staff member at the MIT-IBM Watson AI Lab; William M. Wells, a professor of radiology at MGB and a research scientist in CSAIL; and Tina Kapur, an assistant professor of radiology at MGB. The research will be presented at the International Conference on Learning Representations.

Decoding uncertainty in words

A radiologist writing a report about a chest X-ray might say the image shows a “possible” pneumonia, which is an infection that inflames the air sacs in the lungs. In that case, a doctor could order a follow-up CT scan to confirm the diagnosis.

However, if the radiologist writes that the X-ray shows a “likely” pneumonia, the doctor might begin treatment immediately, such as by prescribing antibiotics, while still ordering additional tests to assess severity.

Trying to measure the calibration, or reliability, of ambiguous natural language terms like “possibly” and “likely” presents many challenges, Wang says.

Existing calibration methods typically rely on the confidence score provided by an AI model, which represents the model’s estimated likelihood that its prediction is correct.

For instance, a weather app might predict an 83 percent chance of rain tomorrow. That model is well-calibrated if, across all instances where it predicts an 83 percent chance of rain, it rains approximately 83 percent of the time.

“But humans use natural language, and if we map these phrases to a single number, it is not an accurate description of the real world. If a person says an event is ‘likely,’ they aren’t necessarily thinking of the exact probability, such as 75 percent,” Wang says.

Rather than trying to map certainty phrases to a single percentage, the researchers’ approach treats them as probability distributions. A distribution describes the range of possible values and their likelihoods — think of the classic bell curve in statistics.

“This captures more nuances of what each word means,” Wang adds.

Assessing and improving calibration

The researchers leveraged prior work that surveyed radiologists to obtain probability distributions that correspond to each diagnostic certainty phrase, ranging from “very likely” to “consistent with.”

For instance, since more radiologists believe the phrase “consistent with” means a pathology is present in a medical image, its probability distribution climbs sharply to a high peak, with most values clustered around the 90 to 100 percent range.

In contrast the phrase “may represent” conveys greater uncertainty, leading to a broader, bell-shaped distribution centered around 50 percent.

Typical methods evaluate calibration by comparing how well a model’s predicted probability scores align with the actual number of positive results.

The researchers’ approach follows the same general framework but extends it to account for the fact that certainty phrases represent probability distributions rather than probabilities.

To improve calibration, the researchers formulated and solved an optimization problem that adjusts how often certain phrases are used, to better align confidence with reality.

They derived a calibration map that suggests certainty terms a radiologist should use to make the reports more accurate for a specific pathology.

“Perhaps, for this dataset, if every time the radiologist said pneumonia was ‘present,’ they changed the phrase to ‘likely present’ instead, then they would become better calibrated,” Wang explains.

When the researchers used their framework to evaluate clinical reports, they found that radiologists were generally underconfident when diagnosing common conditions like atelectasis, but overconfident with more ambiguous conditions like infection.

In addition, the researchers evaluated the reliability of language models using their method, providing a more nuanced representation of confidence than classical methods that rely on confidence scores. 

“A lot of times, these models use phrases like ‘certainly.’ But because they are so confident in their answers, it does not encourage people to verify the correctness of the statements themselves,” Wang adds.

In the future, the researchers plan to continue collaborating with clinicians in the hopes of improving diagnoses and treatment. They are working to expand their study to include data from abdominal CT scans.

In addition, they are interested in studying how receptive radiologists are to calibration-improving suggestions and whether they can mentally adjust their use of certainty phrases effectively.

“Expression of diagnostic certainty is a crucial aspect of the radiology report, as it influences significant management decisions. This study takes a novel approach to analyzing and calibrating how radiologists express diagnostic certainty in chest X-ray reports, offering feedback on term usage and associated outcomes,” says Atul B. Shinagare, associate professor of radiology at Harvard Medical School, who was not involved with this work. “This approach has the potential to improve radiologists’ accuracy and communication, which will help improve patient care.”

The work was funded, in part, by a Takeda Fellowship, the MIT-IBM Watson AI Lab, the MIT CSAIL Wistrom Program, and the MIT Jameel Clinic.

The process of catalysis — in which a material speeds up a chemical reaction — is crucial to the production of many of the chemicals used in our everyday lives. But even though these catalytic processes are widespread, researchers often lack a clear understanding of exactly how they work.

A new analysis by researchers at MIT has shown that an important industrial synthesis process, the production of vinyl acetate, requires a catalyst to take two different forms, which cycle back and forth from one to the other as the chemical process unfolds.

Previously, it had been thought that only one of the two forms was needed. The new findings are published today in the journal Science, in a paper by MIT graduate students Deiaa Harraz and Kunal Lodaya, Bryan Tang PhD ’23, and MIT professor of chemistry and chemical engineering Yogesh Surendranath.

There are two broad classes of catalysts: homogeneous catalysts, which consist of dissolved molecules, and heterogeneous catalysts, which are solid materials whose surface provides the site for the chemical reaction. “For the longest time,” Surendranath says, “there’s been a general view that you either have catalysis happening on these surfaces, or you have them happening on these soluble molecules.” But the new research shows that in the case of vinyl acetate — an important material that goes into many polymer products such as the rubber in the soles of your shoes — there is an interplay between both classes of catalysis.

“What we discovered,” Surendranath explains, “is that you actually have these solid metal materials converting into molecules, and then converting back into materials, in a cyclic dance.”

He adds: “This work calls into question this paradigm where there’s either one flavor of catalysis or another. Really, there could be an interplay between both of them in certain cases, and that could be really advantageous for having a process that’s selective and efficient.”

The synthesis of vinyl acetate has been a large-scale industrial reaction since the 1960s, and it has been well-researched and refined over the years to improve efficiency. This has happened largely through a trial-and-error approach, without a precise understanding of the underlying mechanisms, the researchers say.

While chemists are often more familiar with homogeneous catalysis mechanisms, and chemical engineers are often more familiar with surface catalysis mechanisms, fewer researchers study both. This is perhaps part of the reason that the full complexity of this reaction was not previously captured. But Harraz says he and his colleagues are working at the interface between disciplines. “We’ve been able to appreciate both sides of this reaction and find that both types of catalysis are critical,” he says.

The reaction that produces vinyl acetate requires something to activate the oxygen molecules that are one of the constituents of the reaction, and something else to activate the other ingredients, acetic acid and ethylene. The researchers found that the form of the catalyst that worked best for one part of the process was not the best for the other. It turns out that the molecular form of the catalyst does the key chemistry with the ethylene and the acetic acid, while it’s the surface that ends up doing the activation of the oxygen.

They found that the underlying process involved in interconverting the two forms of the catalyst is actually corrosion, similar to the process of rusting. “It turns out that in rusting, you actually go through a soluble molecular species somewhere in the sequence,” Surendranath says.

The team borrowed techniques traditionally used in corrosion research to study the process. They used electrochemical tools to study the reaction, even though the overall reaction does not require a supply of electricity. By making potential measurements, the researchers determined that the corrosion of the palladium catalyst material to soluble palladium ions is driven by an electrochemical reaction with the oxygen, converting it to water. Corrosion is “one of the oldest topics in electrochemistry,” says Lodaya, “but applying the science of corrosion to understand catalysis is much newer, and was essential to our findings.”

By correlating measurements of catalyst corrosion with other measurements of the chemical reaction taking place, the researchers proposed that it was the corrosion rate that was limiting the overall reaction. “That’s the choke point that’s controlling the rate of the overall process,” Surendranath says.

The interplay between the two types of catalysis works efficiently and selectively “because it actually uses the synergy of a material surface doing what it’s good at and a molecule doing what it’s good at,” Surendranath says. The finding suggests that, when designing new catalysts, rather than focusing on either solid materials or soluble molecules alone, researchers should think about how the interplay of both may open up new approaches.

“Now, with an improved understanding of what makes this catalyst so effective, you can try to design specific materials or specific interfaces that promote the desired chemistry,” Harraz says. Since this process has been worked on for so long, these findings may not necessarily lead to improvements in this specific process of making vinyl acetate, but it does provide a better understanding of why the materials work as they do, and could lead to improvements in other catalytic processes.

Understanding that “catalysts can transit between molecule and material and back, and the role that electrochemistry plays in those transformations, is a concept that we are really excited to expand on,” Lodaya says.

Harraz adds: “With this new understanding that both types of catalysis could play a role, what other catalytic processes are out there that actually involve both? Maybe those have a lot of room for improvement that could benefit from this understanding.”

This work is “illuminating, something that will be worth teaching at the undergraduate level," says Christophe Coperet, a professor of inorganic chemistry at ETH Zurich, who was not associated with the research. “The work highlights new ways of thinking. ... [It] is notable in the sense that it not only reconciles homogeneous and heterogeneous catalysis, but it describes these complex processes as half reactions, where electron transfers can cycle between distinct entities.”

The research was supported, in part, by the National Science Foundation as a Phase I Center for Chemical Innovation; the Center for Interfacial Ionics; and the Gordon and Betty Moore Foundation.

Polymer-coated nanoparticles loaded with therapeutic drugs show significant promise for cancer treatment, including ovarian cancer. These particles can be targeted directly to tumors, where they release their payload while avoiding many of the side effects of traditional chemotherapy.

Over the past decade, MIT Institute Professor Paula Hammond and her students have created a variety of these particles using a technique known as layer-by-layer assembly. They’ve shown that the particles can effectively combat cancer in mouse studies.

To help move these nanoparticles closer to human use, the researchers have now come up with a manufacturing technique that allows them to generate larger quantities of the particles, in a fraction of the time.

“There’s a lot of promise with the nanoparticle systems we’ve been developing, and we’ve been really excited more recently with the successes that we’ve been seeing in animal models for our treatments for ovarian cancer in particular,” says Hammond, who is also MIT’s vice provost for faculty and a member of the Koch Institute for Integrative Cancer Research. “Ultimately, we need to be able to bring this to a scale where a company is able to manufacture these on a large level.”

Hammond and Darrell Irvine, a professor of immunology and microbiology at the Scripps Research Institute, are the senior authors of the new study, which appears today in Advanced Functional Materials. Ivan Pires PhD ’24, now a postdoc at Brigham and Women’s Hospital and a visiting scientist at the Koch Institute, and Ezra Gordon ’24 are the lead authors of paper. Heikyung Suh, an MIT research technician, is also an author.

A streamlined process

More than a decade ago, Hammond’s lab developed a novel technique for building nanoparticles with highly controlled architectures. This approach allows layers with different properties to be laid down on the surface of a nanoparticle by alternately exposing the surface to positively and negatively charged polymers.

Each layer can be embedded with drug molecules or other therapeutics. The layers can also carry targeting molecules that help the particles find and enter cancer cells.

Using the strategy that Hammond’s lab originally developed, one layer is applied at a time, and after each application, the particles go through a centrifugation step to remove any excess polymer. This is time-intensive and would be difficult to scale up to large-scale production, the researchers say.

More recently, a graduate student in Hammond’s lab developed an alternative approach to purifying the particles, known as tangential flow filtration. However, while this streamlined the process, it still was limited by its manufacturing complexity and maximum scale of production.

“Although the use of tangential flow filtration is helpful, it’s still a very small-batch process, and a clinical investigation requires that we would have many doses available for a significant number of patients,” Hammond says.

To create a larger-scale manufacturing method, the researchers used a microfluidic mixing device that allows them to sequentially add new polymer layers as the particles flow through a microchannel within the device. For each layer, the researchers can calculate exactly how much polymer is needed, which eliminates the need to purify the particles after each addition.

“That is really important because separations are the most costly and time-consuming steps in these kinds of systems,” Hammond says.

This strategy eliminates the need for manual polymer mixing, streamlines production, and integrates good manufacturing practice (GMP)-compliant processes. The FDA’s GMP requirements ensure that products meet safety standards and can be manufactured in a consistent fashion, which would be highly challenging and costly using the previous step-wise batch process. The microfluidic device that the researchers used in this study is already used for GMP manufacturing of other types of nanoparticles, including mRNA vaccines.

“With the new approach, there’s much less chance of any sort of operator mistake or mishaps,” Pires says. “This is a process that can be readily implemented in GMP, and that’s really the key step here. We can create an innovation within the layer-by-layer nanoparticles and quickly produce it in a manner that we could go into clinical trials with.”

Scaled-up production

Using this approach, the researchers can generate 15 milligrams of nanoparticles (enough for about 50 doses) in just a few minutes, while the original technique would take close to an hour to create the same amount. This could enable the production of more than enough particles for clinical trials and patient use, the researchers say.

“To scale up with this system, you just keep running the chip, and it is much easier to produce more of your material,” Pires says.

To demonstrate their new production technique, the researchers created nanoparticles coated with a cytokine called interleukin-12 (IL-12). Hammond’s lab has previously shown that IL-12 delivered by layer-by-layer nanoparticles can activate key immune cells and slow ovarian tumor growth in mice.

In this study, the researchers found that IL-12-loaded particles manufactured using the new technique showed similar performance as the original layer-by-layer nanoparticles. And, not only do these nanoparticles bind to cancer tissue, but they show a unique ability to not enter the cancer cells. This allows the nanoparticles to serve as markers on the cancer cells that activate the immune system locally in the tumor. In mouse models of ovarian cancer, this treatment can lead to both tumor growth delay and even cures.

The researchers have filed for a patent on the technology and are now working with MIT’s Deshpande Center for Technological Innovation in hopes of potentially forming a company to commercialize the technology. While they are initially focusing on cancers of the abdominal cavity, such as ovarian cancer, the work could also be applied to other types of cancer, including glioblastoma, the researchers say.

The research was funded by the U.S. National Institutes of Health, the Marble Center for Nanomedicine, the Deshpande Center for Technological Innovation, and the Koch Institute Support (core) Grant from the National Cancer Institute.

When major disasters hit and structures collapse, people can become trapped under rubble. Extricating victims from these hazardous environments can be dangerous and physically exhausting. To help rescue teams navigate these structures, MIT Lincoln Laboratory, in collaboration with researchers at the University of Notre Dame, developed the Soft Pathfinding Robotic Observation Unit (SPROUT). SPROUT is a vine robot — a soft robot that can grow and maneuver around obstacles and through small spaces. First responders can deploy SPROUT under collapsed structures to explore, map, and find optimum ingress routes through debris. 

"The urban search-and-rescue environment can be brutal and unforgiving, where even the most hardened technology struggles to operate. The fundamental way a vine robot works mitigates a lot of the challenges that other platforms face," says Chad Council, a member of the SPROUT team, which is led by Nathaniel Hanson. The program is conducted out of the laboratory's Human Resilience Technology Group. 

First responders regularly integrate technology, such as cameras and sensors, into their workflows to understand complex operating environments. However, many of these technologies have limitations. For example, cameras specially built for search-and-rescue operations can only probe on a straight path inside of a collapsed structure. If a team wants to search further into a pile, they need to cut an access hole to get to the next area of the space. Robots are good for exploring on top of rubble piles, but are ill-suited for searching in tight, unstable structures and costly to repair if damaged. The challenge that SPROUT addresses is how to get under collapsed structures using a low-cost, easy-to-operate robot that can carry cameras and sensors and traverse winding paths. 

SPROUT is composed of an inflatable tube made of airtight fabric that unfurls from a fixed base. The tube inflates with air, and a motor controls its deployment. As the tube extends into rubble, it can flex around corners and squeeze through narrow passages. A camera and other sensors mounted to the tip of the tube image and map the environment the robot is navigating. An operator steers SPROUT with joysticks, watching a screen that displays the robot's camera feed. Currently, SPROUT can deploy up to 10 feet, and the team is working on expanding it to 25 feet.

When building SPROUT, the team overcame a number of challenges related to the robot's flexibility. Because the robot is made of a deformable material that bends at many points, determining and controlling the robot's shape as it unfurls through the environment is difficult — think of trying to control an expanding wiggly sprinkler toy. Pinpointing how to apply air pressure within the robot so that steering is as simple as pointing the joystick forward to make the robot move forward was essential for system adoption by emergency responders. In addition, the team had to design the tube to minimize friction while the robot grows and engineer the controls for steering.

While a teleoperated system is a good starting point for assessing the hazards of void spaces, the team is also finding new ways to apply robot technologies to the domain, such as using data captured by the robot to build maps of the subsurface voids. "Collapse events are rare but devastating events. In robotics, we would typically want ground truth measurements to validate our approaches, but those simply don't exist for collapsed structures," Hanson says. To solve this problem, Hanson and his team made a simulator that allows them to create realistic depictions of collapsed structures and develop algorithms that map void spaces.

SPROUT was developed in collaboration with Margaret Coad, a professor at the University of Notre Dame and an MIT graduate. When looking for collaborators, Hanson — a graduate of Notre Dame — was already aware of Coad's work on vine robots for industrial inspection. Coad's expertise, together with the laboratory's experience in engineering, strong partnership with urban search-and-rescue teams, and ability to develop fundamental technologies and prepare them for  transition to industry, "made this a really natural pairing to join forces and work on research for a traditionally underserved community," Hanson says. "As one of the primary inventors of vine robots, Professor Coad brings invaluable expertise on the fabrication and modeling of these robots."

Lincoln Laboratory tested SPROUT with first responders at the  Massachusetts Task Force 1  training site in Beverly, Massachusetts. The tests allowed the researchers to improve the durability and portability of the robot and learn how to grow and steer the robot more efficiently. The team is planning a larger field study this spring.

"Urban search-and-rescue teams and first responders serve critical roles in their communities but typically have little-to-no research and development budgets," Hanson says. "This program has enabled us to push the technology readiness level of vine robots to a point where responders can engage with a hands-on demonstration of the system."

Sensing in constrained spaces is not a problem unique to disaster response communities, Hanson adds. The team envisions the technology being used in the maintenance of military systems or critical infrastructure with difficult-to-access locations.

The initial program focused on mapping void spaces, but future work aims to localize hazards and assess the viability and safety of operations through rubble. "The mechanical performance of the robots has an immediate effect, but the real goal is to rethink the way sensors are used to enhance situational awareness for rescue teams," says Hanson. "Ultimately, we want SPROUT to provide a complete operating picture to teams before anyone enters a rubble pile." 

Imagine a coffee company trying to optimize its supply chain. The company sources beans from three suppliers, roasts them at two facilities into either dark or light coffee, and then ships the roasted coffee to three retail locations. The suppliers have different fixed capacity, and roasting costs and shipping costs vary from place to place.

The company seeks to minimize costs while meeting a 23 percent increase in demand.

Wouldn’t it be easier for the company to just ask ChatGPT to come up with an optimal plan? In fact, for all their incredible capabilities, large language models (LLMs) often perform poorly when tasked with directly solving such complicated planning problems on their own.

Rather than trying to change the model to make an LLM a better planner, MIT researchers took a different approach. They introduced a framework that guides an LLM to break down the problem like a human would, and then automatically solve it using a powerful software tool.

A user only needs to describe the problem in natural language — no task-specific examples are needed to train or prompt the LLM. The model encodes a user’s text prompt into a format that can be unraveled by an optimization solver designed to efficiently crack extremely tough planning challenges.

During the formulation process, the LLM checks its work at multiple intermediate steps to make sure the plan is described correctly to the solver. If it spots an error, rather than giving up, the LLM tries to fix the broken part of the formulation.

When the researchers tested their framework on nine complex challenges, such as minimizing the distance warehouse robots must travel to complete tasks, it achieved an 85 percent success rate, whereas the best baseline only achieved a 39 percent success rate.

The versatile framework could be applied to a range of multistep planning tasks, such as scheduling airline crews or managing machine time in a factory.

“Our research introduces a framework that essentially acts as a smart assistant for planning problems. It can figure out the best plan that meets all the needs you have, even if the rules are complicated or unusual,” says Yilun Hao, a graduate student in the MIT Laboratory for Information and Decision Systems (LIDS) and lead author of a paper on this research.

She is joined on the paper by Yang Zhang, a research scientist at the MIT-IBM Watson AI Lab; and senior author Chuchu Fan, an associate professor of aeronautics and astronautics and LIDS principal investigator. The research will be presented at the International Conference on Learning Representations.

Optimization 101

The Fan group develops algorithms that automatically solve what are known as combinatorial optimization problems. These vast problems have many interrelated decision variables, each with multiple options that rapidly add up to billions of potential choices.

Humans solve such problems by narrowing them down to a few options and then determining which one leads to the best overall plan. The researchers’ algorithmic solvers apply the same principles to optimization problems that are far too complex for a human to crack.

But the solvers they develop tend to have steep learning curves and are typically only used by experts.

“We thought that LLMs could allow nonexperts to use these solving algorithms. In our lab, we take a domain expert’s problem and formalize it into a problem our solver can solve. Could we teach an LLM to do the same thing?” Fan says.

Using the framework the researchers developed, called LLM-Based Formalized Programming (LLMFP), a person provides a natural language description of the problem, background information on the task, and a query that describes their goal.

Then LLMFP prompts an LLM to reason about the problem and determine the decision variables and key constraints that will shape the optimal solution.

LLMFP asks the LLM to detail the requirements of each variable before encoding the information into a mathematical formulation of an optimization problem. It writes code that encodes the problem and calls the attached optimization solver, which arrives at an ideal solution.

“It is similar to how we teach undergrads about optimization problems at MIT. We don’t teach them just one domain. We teach them the methodology,” Fan adds.

As long as the inputs to the solver are correct, it will give the right answer. Any mistakes in the solution come from errors in the formulation process.

To ensure it has found a working plan, LLMFP analyzes the solution and modifies any incorrect steps in the problem formulation. Once the plan passes this self-assessment, the solution is described to the user in natural language.

Perfecting the plan

This self-assessment module also allows the LLM to add any implicit constraints it missed the first time around, Hao says.

For instance, if the framework is optimizing a supply chain to minimize costs for a coffeeshop, a human knows the coffeeshop can’t ship a negative amount of roasted beans, but an LLM might not realize that.

The self-assessment step would flag that error and prompt the model to fix it.

“Plus, an LLM can adapt to the preferences of the user. If the model realizes a particular user does not like to change the time or budget of their travel plans, it can suggest changing things that fit the user’s needs,” Fan says.

In a series of tests, their framework achieved an average success rate between 83 and 87 percent across nine diverse planning problems using several LLMs. While some baseline models were better at certain problems, LLMFP achieved an overall success rate about twice as high as the baseline techniques.

Unlike these other approaches, LLMFP does not require domain-specific examples for training. It can find the optimal solution to a planning problem right out of the box.

In addition, the user can adapt LLMFP for different optimization solvers by adjusting the prompts fed to the LLM.

“With LLMs, we have an opportunity to create an interface that allows people to use tools from other domains to solve problems in ways they might not have been thinking about before,” Fan says.

In the future, the researchers want to enable LLMFP to take images as input to supplement the descriptions of a planning problem. This would help the framework solve tasks that are particularly hard to fully describe with natural language.

This work was funded, in part, by the Office of Naval Research and the MIT-IBM Watson AI Lab.

How far would you go for a good meal? For some of the ocean’s top predators, maintaining a decent diet requires some surprisingly long-distance dives.

MIT oceanographers have found that big fish like tuna and swordfish get a large fraction of their food from the ocean’s twilight zone — a cold and dark layer of the ocean about half a mile below the surface, where sunlight rarely penetrates. Tuna and swordfish have been known to take extreme plunges, but it was unclear whether these deep dives were for food, and to what extent the fishes’ diet depends on prey in the twilight zone.

In a study published recently in the ICES Journal of Marine Science, the MIT student-led team reports that the twilight zone is a major food destination for three predatory fish — bigeye tuna, yellowfin tuna, and swordfish. While the three species swim primarily in the shallow open ocean, the scientists found these fish are sourcing between 50 and 60 percent of their diet from the twilight zone.

The findings suggest that tuna and swordfish rely more heavily on the twilight zone than scientists had assumed. This implies that any change to the twilight zone’s food web, such as through increased fishing, could negatively impact fisheries of more shallow tuna and swordfish.

“There is increasing interest in commercial fishing in the ocean’s twilight zone,” says Ciara Willis, the study’s lead author, who was a PhD student in the MIT-Woods Hole Oceanographic Institution (WHOI) Joint Program when conducting the research and is now a postdoc at WHOI. “If we start heavily fishing that layer of the ocean, our study suggests that could have profound implications for tuna and swordfish, which are very reliant on the twilight zone and are highly valuable existing fisheries.”

The study’s co-authors include Kayla Gardener of MIT-WHOI, and WHOI researchers Martin Arostegui, Camrin Braun, Leah Hougton, Joel Llopiz, Annette Govindarajan, and Simon Thorrold, along with Walt Golet at the University of Maine.

Deep-ocean buffet

The ocean’s twilight zone is a vast and dim layer that lies between the sunlit surface waters and the ocean’s permanently dark, midnight zone. Also known as the midwater, or mesopelagic layer, the twilight zone stretches between 200 and 1,000 meters below the ocean’s surface and is home to a huge variety of organisms that have adapted to live in the darkness.

“This is a really understudied region of the ocean, and it’s filled with all these fantastic, weird animals,” Willis says.

In fact, it’s estimated that the biomass of fish in the twilight zone is somewhere close to 10 billion tons, much of which is concentrated in layers at certain depths. By comparison, the marine life that lives closer to the surface, Willis says, is “a thin soup,” which is slim pickings for large predators.

“It’s important for predators in the open ocean to find concentrated layers of food. And I think that’s what drives them to be interested in the ocean’s twilight zone,” Willis says. “We call it the ‘deep ocean buffet.’”

And much of this buffet is on the move. Many kinds of fish, squid, and other deep-sea organisms in the twilight zone will swim up to the surface each night to find food. This twilight community will descend back into darkness at dawn to avoid detection.

Scientists have observed that many large predatory fish will make regular dives into the twilight zone, presumably to feast on the deep-sea bounty. For instance, bigeye tuna spend much of their day making multiple short, quick plunges into the twilight zone, while yellowfin tuna dive down every few days to weeks. Swordfish, in contrast, appear to follow the daily twilight migration, feeding on the community as it rises and falls each day.

“We’ve known for a long time that these fish and many other predators feed on twilight zone prey,” Willis says. “But the extent to which they rely on this deep-sea food web for their forage has been unclear.”

Twilight signal

For years, scientists and fishers have found remnants of fish from the twilight zone in the stomach contents of larger, surface-based predators. This suggests that predator fish do indeed feed on twilight food, such as lanternfish, certain types of squid, and long, snake-like fish called barracudina. But, as Willis notes, stomach contents give just a “snapshot” of what a fish ate that day.

She and her colleagues wanted to know how big a role twilight food plays in the general diet of predator fish. For their new study, the team collaborated with fishermen in New Jersey and Florida, who fish for a living in the open ocean. They supplied the team with small tissue samples of their commercial catch, including samples of bigeye tuna, yellowfin tuna, and swordfish.

Willis and her advisor, Senior Scientist Simon Thorrold, brought the samples back to Thorrold’s lab at WHOI and analyzed the fish bits for essential amino acids — the key building blocks of proteins. Essential amino acids are only made by primary producers, or members of the base of the food web, such as phytoplankton, microbes, and fungi. Each of these producers makes essential amino acids with a slightly different carbon isotope configuration that then is conserved as the producers are consumed on up their respective food chains.

“One of the hypotheses we had was that we’d be able to distinguish the carbon isotopic signature of the shallow ocean, which would logically be more phytoplankton-based, versus the deep ocean, which is more microbially based,” Willis says.

The researchers figured that if a fish sample had one carbon isotopic make-up over another, it would be a sign that that fish feeds more on food from the deep, rather than shallow waters.

“We can use this [carbon isotope signature] to infer a lot about what food webs they’ve been feeding in, over the last five to eight months,” Willis says.

The team looked at carbon isotopes in tissue samples from over 120 samples including bigeye tuna, yellowfin tuna, and swordfish. They found that individuals from all three species contained a substantial amount of carbon derived from sources in the twilight zone. The researchers estimate that, on average, food from the twilight zone makes up 50 to 60 percent of the diet of the three predator species, with some slight variations among species.

“We saw the bigeye tuna were far and away the most consistent in where they got their food from. They didn’t vary much from individual to individual,” Willis says. “Whereas the swordfish and yellowfin tuna were more variable. That means if you start having big-scale fishing in the twilight zone, the bigeye tuna might be the ones who are most at risk from food web effects.”

The researchers note there has been increased interest in commercially fishing the twilight zone. While many fish in that region are not edible for humans, they are starting to be harvested as fishmeal and fish oil products. In ongoing work, Willis and her colleagues are evaluating the potential impacts to tuna fisheries if the twilight zone becomes a target for large-scale fishing.

“If predatory fish like tunas have 50 percent reliance on twilight zone food webs, and we start heavily fishing that region, that could lead to uncertainty around the profitability of tuna fisheries,” Willis says. “So we need to be very cautious about impacts on the twilight zone and the larger ocean ecosystem.”

This work was part of the Woods Hole Oceanographic Institution’s Ocean Twilight Zone Project, funded as part of the Audacious Project housed at TED. Willis was additionally supported by the Natural Sciences and Engineering Research Council of Canada and the MIT Martin Family Society of Fellows for Sustainability.

It’s tricky to predict precisely what the impacts of climate change will be, given the many variables involved. To predict the impacts of a warmer world on plant life, some researchers look at urban “heat islands,” where, because of the effects of urban structures, temperatures consistently run a few degrees higher than those of the surrounding rural areas. This enables side-by-side comparisons of plant responses.

But a new study by researchers at MIT and Harvard University has found that, at least for forests, urban heat islands are a poor proxy for global warming, and this may have led researchers to underestimate the impacts of warming in some cases. The discrepancy, they found, has a lot to do with the limited genetic diversity of urban tree species.

The findings appear in the journal PNAS, in a paper by MIT postdoc Meghan Blumstein, professor of civil and environmental engineering David Des Marais, and four others.

“The appeal of these urban temperature gradients is, well, it’s already there,” says Des Marais. “We can’t look into the future, so why don’t we look across space, comparing rural and urban areas?” Because such data is easily obtainable, methods comparing the growth of plants in cities with similar plants outside them have been widely used, he says, and have been quite useful. Researchers did recognize some shortcomings to this approach, including significant differences in availability of some nutrients such as nitrogen. Still, “a lot of ecologists recognized that they weren’t perfect, but it was what we had,” he says.

Most of the research by Des Marais’ group is lab-based, under conditions tightly controlled for temperature, humidity, and carbon dioxide concentration. While there are a handful of experimental sites where conditions are modified out in the field, for example using heaters around one or a few trees, “those are super small-scale,” he says. “When you’re looking at these longer-term trends that are occurring over space that’s quite a bit larger than you could reasonably manipulate, an important question is, how do you control the variables?”

Temperature gradients have offered one approach to this problem, but Des Marais and his students have also been focusing on the genetics of the tree species involved, comparing those sampled in cities to the same species sampled in a natural forest nearby. And it turned out there were differences, even between trees that appeared similar.

“So, lo and behold, you think you’re only letting one variable change in your model, which is the temperature difference from an urban to a rural setting,” he says, “but in fact, it looks like there was also a genotypic diversity that was not being accounted for.”

The genetic differences meant that the plants being studied were not representative of those in the natural environment, and the researchers found that the difference was actually masking the impact of warming. The urban trees, they found, were less affected than their natural counterparts in terms of when the plants’ leaves grew and unfurled, or “leafed out,” in the spring.

The project began during the pandemic lockdown, when Blumstein was a graduate student. She had a grant to study red oak genotypes across New England, but was unable to travel because of lockdowns. So, she concentrated on trees that were within reach in Cambridge, Massachusetts. She then collaborated with people doing research at the Harvard Forest, a research forest in rural central Massachusetts. They collected three years of data from both locations, including the temperature profiles, the leafing-out timing, and the genetic profiles of the trees. Though the study was looking at red oaks specifically, the researchers say the findings are likely to apply to trees broadly.

At the time, researchers had just sequenced the oak tree genome, and that allowed Blumstein and her colleagues to look for subtle differences among the red oaks in the two locations. The differences they found showed that the urban trees were more resistant to the effects of warmer temperatures than were those in the natural environment.

“Initially, we saw these results and we were sort of like, oh, this is a bad thing,” Des Marais says. “Ecologists are getting this heat island effect wrong, which is true.” Fortunately, this can be easily corrected by factoring in genomic data. “It’s not that much more work, because sequencing genomes is so cheap and so straightforward. Now, if someone wants to look at an urban-rural gradient and make these kinds of predictions, well, that’s fine. You just have to add some information about the genomes.”

It's not surprising that this genetic variation exists, he says, since growers have learned by trial and error over the decades which varieties of trees tend to thrive in the difficult urban environment, with typically poor soil, poor drainage, and pollution. “As a result, there’s just not much genetic diversity in our trees within cities.”

The implications could be significant, Des Marais says. When the Intergovernmental Panel on Climate Change (IPCC) releases its regular reports on the status of the climate, “one of the tools the IPCC has to predict future responses to climate change with respect to temperature are these urban-to-rural gradients.” He hopes that these new findings will be incorporated into their next report, which is just being drafted. “If these results are generally true beyond red oaks, this suggests that the urban heat island approach to studying plant response to temperature is underpredicting how strong that response is.”

The research team included Sophie Webster, Robin Hopkins, and David Basler from Harvard University and Jie Yun from MIT. The work was supported by the National Science Foundation, the Bullard Fellowship at the Harvard Forest, and MIT.

Metamaterials are artificially-structured materials with extraordinary properties not easily found in nature. With engineered three-dimensional (3D) geometries at the micro- and nanoscale, these architected materials achieve unique mechanical and physical properties with capabilities beyond those of conventional materials — and have emerged over the past decade as a promising way to engineering challenges where all other existing materials have lacked success.

Architected materials exhibit unique mechanical and functional properties, but their full potential remains untapped due to challenges in design, fabrication, and characterization. Improvements and scalability in this space could help transform a range of industries, from biomedical implants, sports equipment, automotive and aerospace, and energy and electronics.

“Advances in scalable fabrication, high-throughput testing, and AI-driven design optimization could revolutionize the mechanics and materials science disciplines, enabling smarter, more adaptive materials that redefine engineering and everyday technologies,” says Carlos Portela, the Robert N. Noyce Career Development Professor and assistant professor of mechanical engineering at MIT.

In a Perspective published this month in the journal Nature Materials, Portela and James Surjadi, a postdoc in mechanical engineering, discuss key hurdles, opportunities, and future applications in the field of mechanical metamaterials. The paper is titled “Enabling three-dimensional architected materials across length scales and timescales.”

“The future of the field requires innovation in fabricating these materials across length scales, from nano to macro, and progress in understanding them at a variety of time scales, from slow deformation to dynamic impact,” says Portela, adding that it also demands interdisciplinary collaboration.

A Perspective is a peer-reviewed content type that the journal uses to invite reflection or discussion on matters that may be speculative, controversial, or highly technical, and where the subject matter may not meet the criteria for a Review.

“We felt like our field, following substantial progress over the last decade, is still facing two bottlenecks: issues scaling up, and no knowledge or understanding of properties under dynamic conditions,” says Portela, discussing the decision to write the piece.

Portela and Surjadi’s paper summarizes state-of-the-art approaches and highlights existing knowledge gaps in material design, fabrication, and characterization. It also proposes a roadmap to accelerate the discovery of architected materials with programmable properties via the synergistic combination of high-throughput experimentation and computational efforts, toward leveraging emerging artificial intelligence and machine learning techniques for their design and optimization.

“High-throughput miniaturized experiments, non-contact characterization, and benchtop extreme-condition methods will generate rich datasets for the implementation of data-driven models, accelerating the optimization and discovery of metamaterials with unique properties,” says Surjadi.

The Portela Lab’s motto is “architected mechanics and materials across scales.” The Perspective aims to bridge the gap between fundamental research and real-world applications of next-generation architected materials, and it presents a vision the lab has been working toward for the past four years.

Around 11 billion tons of goods, or about 1.5 tons per person worldwide, are transported by sea each year, representing about 90 percent of global trade by volume. Internationally, the merchant shipping fleet numbers around 110,000 vessels. These ships, and the ports that service them, are significant contributors to the local and global economy — and they’re significant contributors to greenhouse gas emissions.

A new consortium, formalized in a signing ceremony at MIT last week, aims to address climate-harming emissions in the maritime shipping industry, while supporting efforts for environmentally friendly operation in compliance with the decarbonization goals set by the International Maritime Organization.

“This is a timely collaboration with key stakeholders from the maritime industry with a very bold and interdisciplinary research agenda that will establish new technologies and evidence-based standards,” says Themis Sapsis, the William Koch Professor of Marine Technology at MIT and the director of MIT’s Center for Ocean Engineering. “It aims to bring the best from MIT in key areas for commercial shipping, such as nuclear technology for commercial settings, autonomous operation and AI methods, improved hydrodynamics and ship design, cybersecurity, and manufacturing.” 

Co-led by Sapsis and Fotini Christia, the Ford International Professor of the Social Sciences; director of the Institute for Data, Systems, and Society (IDSS); and director of the MIT Sociotechnical Systems Research Center, the newly-launched MIT Maritime Consortium (MC) brings together MIT collaborators from across campus, including the Center for Ocean Engineering, which is housed in the Department of Mechanical Engineering; IDSS, which is housed in the MIT Schwarzman College of Computing; the departments of Nuclear Science and Engineering and Civil and Environmental Engineering; MIT Sea Grant; and others, with a national and an international community of industry experts.

The Maritime Consortium’s founding members are the American Bureau of Shipping (ABS), Capital Clean Energy Carriers Corp., and HD Korea Shipbuilding and Offshore Engineering. Innovation members are Foresight-Group, Navios Maritime Partners L.P., Singapore Maritime Institute, and Dorian LPG.

“The challenges the maritime industry faces are challenges that no individual company or organization can address alone,” says Christia. “The solution involves almost every discipline from the School of Engineering, as well as AI and data-driven algorithms, and policy and regulation — it’s a true MIT problem.”

Researchers will explore new designs for nuclear systems consistent with the techno-economic needs and constraints of commercial shipping, economic and environmental feasibility of alternative fuels, new data-driven algorithms and rigorous evaluation criteria for autonomous platforms in the maritime space, cyber-physical situational awareness and anomaly detection, as well as 3D printing technologies for onboard manufacturing. Collaborators will also advise on research priorities toward evidence-based standards related to MIT presidential priorities around climate, sustainability, and AI.

MIT has been a leading center of ship research and design for over a century, and is widely recognized for contributions to hydrodynamics, ship structural mechanics and dynamics, propeller design, and overall ship design, and its unique educational program for U.S. Navy Officers, the Naval Construction and Engineering Program. Research today is at the forefront of ocean science and engineering, with significant efforts in fluid mechanics and hydrodynamics, acoustics, offshore mechanics, marine robotics and sensors, and ocean sensing and forecasting. The consortium’s academic home at MIT also opens the door to cross-departmental collaboration across the Institute.

The MC will launch multiple research projects designed to tackle challenges from a variety of angles, all united by cutting-edge data analysis and computation techniques. Collaborators will research new designs and methods that improve efficiency and reduce greenhouse gas emissions, explore feasibility of alternative fuels, and advance data-driven decision-making, manufacturing and materials, hydrodynamic performance, and cybersecurity.

“This consortium brings a powerful collection of significant companies that, together, has the potential to be a global shipping shaper in itself,” says Christopher J. Wiernicki SM ’85, chair and chief executive officer of ABS. 

“The strength and uniqueness of this consortium is the members, which are all world-class organizations and real difference makers. The ability to harness the members’ experience and know-how, along with MIT’s technology reach, creates real jet fuel to drive progress,” Wiernicki says. “As well as researching key barriers, bottlenecks, and knowledge gaps in the emissions challenge, the consortium looks to enable development of the novel technology and policy innovation that will be key. Long term, the consortium hopes to provide the gravity we will need to bend the curve.”

Bar graphs and other charts provide a simple way to communicate data, but are, by definition, difficult to translate for readers who are blind or low-vision. Designers have developed methods for converting these visuals into “tactile charts,” but guidelines for doing so are extensive (for example, the Braille Authority of North America’s 2022 guidebook is 426 pages long). The process also requires understanding different types of software, as designers often draft their chart in programs like Adobe Illustrator and then translate it into Braille using another application.

Researchers from MIT’s Computer Science and Artificial Intelligence Laboratory (CSAIL) have now developed an approach that streamlines the design process for tactile chart designers. Their program, called “Tactile Vega-Lite,” can take data from something like an Excel spreadsheet and turn it into both a standard visual chart and a touch-based one. Design standards are hardwired as default rules within the program to help educators and designers automatically create accessible tactile charts.

The tool could make it easier for blind and low-vision readers to understand many graphics, such as a bar chart comparing minimum wages across states or a line graph tracking countries’ GDPs over time. To bring your designs to the real world, you can tweak your chart in Tactile Vega-Lite and then send its file to a Braille embosser (which prints text as readable dots).

This spring, the researchers will present Tactile Vega-Lite in a paper at the Association of Computing Machinery Conference on Human Factors in Computing Systems. According to lead author Mengzhu “Katie” Chen SM ’25, the tool strikes a balance between the precision that design professionals want for editing and the efficiency educators need to create tactile charts quickly.

“We interviewed teachers who wanted to make their lessons accessible to blind and low-vision students, and designers experienced in putting together tactile charts,” says Chen, a recent CSAIL affiliate and master's graduate in electrical engineering and computer science and the Program in System Design and Management. “Since their needs differ, we designed a program that’s easy to use, provides instant feedback when you want to make tweaks, and implements accessibility guidelines.”

Data you can feel

The researchers’ program builds off of their 2017 visualization tool Vega-Lite by automatically encoding both a flat, standard chart and a tactile one. Senior author and MIT postdoc Jonathan Zong SM ’20, PhD ’24 points out that the program makes intuitive design decisions so users don’t have to.

“Tactile Vega-Lite has smart defaults to ensure proper spacing, layout, and texture and Braille conversion, following best practices to create good touch-based reading experiences,” says Zong, who is also a fellow at the Berkman Klein Center for Internet and Society at Harvard University and an incoming assistant professor at the University of Colorado. “Building on existing guidelines and our interviews with experts, the goal is for teachers or visual designers without a lot of tactile design expertise to quickly convey data in a clear way for tactile readers to explore and understand.”

Tactile Vega-Lite’s code editor allows users to customize axis labels, tick marks, and other elements. Different features within the chart are represented by abstractions — or summaries of a longer body of code — that can be modified. These shortcuts allow you to write brief phrases that tweak the design of your chart. For example, if you want to change how the bars in your graph are filled out, you could change the code in the “Texture” section from “dottedFill” to “verticalFill” to replace small circles with upward lines.

To understand how these abstractions work, the researchers added a gallery of examples. Each one includes a phrase and what change that code leads to. Still, the team is looking to refine Tactile Vega-Lite’s user interface to make it more accessible to users less familiar with coding. Instead of using abstractions for edits, you could click on different buttons.

Chen says she and her colleagues are hoping to add machine-specific customizations to their program. This would allow users to preview how their tactile chart would look before it’s fabricated by an embossing machine and make edits according to the device’s specifications.

While Tactile Vega-Lite can streamline the many steps it usually takes to make a tactile chart, Zong emphasizes that it doesn’t replace an expert doing a final check-over for guideline compliance. The researchers are continuing to incorporate Braille design rules into their program, but caution that human review will likely remain the best practice.

“The ability to design tactile graphics efficiently, particularly without specialized software, is important for providing equal access of information to tactile readers,” says Stacy Fontenot, owner of Font to Dot, who wasn’t involved in the research. “Graphics that follow current guidelines and standards are beneficial for the reader as consistency is paramount, especially with complex, data-filled graphics. Tactile Vega-Lite has a straightforward interface for creating informative tactile graphics quickly and accurately, thereby reducing the design time in providing quality graphics to tactile readers.”

Chen and Zong wrote the paper with Isabella Pineros ’23, MEng ’24 and MIT Associate Professor Arvind Satyanarayan. The researchers’ work was supported by a National Science Foundation grant.

The CSAIL team also incorporated input from Rich Caloggero from MIT’s Disability and Access Services, as well as the Lighthouse for the Blind, which let them observe technical design workflows as part of the project.

Reducing the amount of agricultural sprays used by farmers — including fertilizers, pesticides and herbicides — could cut down the amount of polluting runoff that ends up in the environment while at the same time reducing farmers’ costs and perhaps even enhancing their productivity. A classic win-win-win.

A team of researchers at MIT and a spinoff company they launched has developed a system to do just that. Their technology adds a thin coating around droplets as they are being sprayed onto a field, greatly reducing their tendency to bounce off leaves and end up wasted on the ground. Instead, the coated droplets stick to the leaves as intended.

The research is described today in the journal Soft Matter, in a paper by recent MIT alumni Vishnu Jayaprakash PhD ’22 and Sreedath Panat PhD ’23, graduate student Simon Rufer, and MIT professor of mechanical engineering Kripa Varanasi.

A recent study found that if farmers didn’t use pesticides, they would lose 78 percent of fruit, 54 percent of vegetable, and 32 percent of cereal production. Despite their importance, a lack of technology that monitors and optimizes sprays has forced farmers to rely on personal experience and rules of thumb to decide how to apply these chemicals. As a result, these chemicals tend to be over-sprayed, leading to runoff and chemicals ending up in waterways or building up in the soil.

Pesticides take a significant toll on global health and the environment, the researchers point out. A recent study found that 31 percent of agricultural soils around the world were at high risk from pesticide pollution. And agricultural chemicals are a major expense for farmers: In the U.S., they spend $16 billion a year just on pesticides.

Making spraying more efficient is one of the best ways to make food production more sustainable and economical. Agricultural spraying essentially boils down to mixing chemicals into water and spraying water droplets onto plant leaves, which are often inherently water-repellent. “Over more than a decade of research in my lab at MIT, we have developed fundamental understandings of spraying and the interaction between droplets and plants — studying when they bounce and all the ways we have to make them stick better and enhance coverage,” Varanasi says.

The team had previously found a way to reduce the amount of sprayed liquid that bounces away from the leaves it strikes, which involved using two spray nozzles instead of one and spraying mixtures with opposite electrical charges. But they found that farmers were reluctant to take on the expense and effort of converting their spraying equipment to a two-nozzle system. So, the team looked for a simpler alternative.

They discovered they could achieve the same improvement in droplet retention using a single-nozzle system that can be easily adapted to existing sprayers. Instead of giving the droplets of pesticide an electric charge, they coat each droplet with a vanishingly thin layer of an oily material.

In their new study, they conducted lab experiments with high-speed cameras. When they sprayed droplets with no special treatment onto a water-repelling (hydrophobic) surface similar to that of many plant leaves, the droplets initially spread out into a pancake-like disk, then rebounded back into a ball and bounced away. But when the researchers coated the surface of the droplets with a tiny amount of oil — making up less than 1 percent of the droplet’s liquid — the droplets spread out and then stayed put. The treatment improved the droplets’ “stickiness” by as much as a hundredfold.

“When these droplets are hitting the surface and as they expand, they form this oil ring that essentially pins the droplet to the surface,” Rufer says. The researchers tried a wide variety of conditions, he says, explaining that they conducted hundreds of experiments, “with different impact velocities, different droplet sizes, different angles of inclination, all the things that fully characterize this phenomenon.” Though different oils varied in their effectiveness, all of them were effective. “Regardless of the impact velocity and the oils, we saw that the rebound height was significantly lower,” he says.

The effect works with remarkably small amounts of oil. In their initial tests they used 1 percent oil compared to the water, then they tried a 0.1 percent, and even .01. The improvement in droplets sticking to the surface continued at a 0.1 percent, but began to break down beyond that. “Basically, this oil film acts as a way to trap that droplet on the surface, because oil is very attracted to the surface and sort of holds the water in place,” Rufer says.

In the researchers’ initial tests they used soybean oil for the coating, figuring this would be a familiar material for the farmers they were working with, many of whom were growing soybeans. But it turned out that though they were producing the beans, the oil was not part of their usual supply chain for use on the farm. In further tests, the researchers found that several chemicals that farmers were already routinely using in their spraying, called surfactants and adjuvants, could be used instead, and that some of these provided the same benefits in keeping the droplets stuck on the leaves.

“That way,” Varanasi says, “we’re not introducing a new chemical or changed chemistries into their field, but they’re using things they’ve known for a long time.”

Varanasi and Jayaprakash formed a company called AgZen to commercialize the system. In order to prove how much their coating system improves the amount of spray that stays on the plant, they first had to develop a system to monitor spraying in real time. That system, which they call RealCoverage, has been deployed on farms ranging in size from a few dozen acres to hundreds of thousands of acres, and many different crop types, and has saved farmers 30 to 50 percent on their pesticide expenditures, just by improving the controls on the existing sprays. That system is being deployed to 920,000 acres of crops in 2025, the company says, including some in California, Texas, the Midwest, France and Italy. Adding the cloaking system using new nozzles, the researchers say, should yield at least another doubling of efficiency.

“You could give back a billion dollars to U.S. growers if you just saved 6 percent of their pesticide budget,” says Jayaprakash, lead author of the research paper and CEO of AgZen. “In the lab we got 300 percent of extra product on the plant. So that means we could get orders of magnitude reductions in the amount of pesticides that farmers are spraying.”

Farmers had already been using these surfactant and adjuvant chemicals as a way to enhance spraying effectiveness, but they were mixing it with a water solution. For it to have any effect, they had to use much more of these materials, risking causing burns to the plants. The new coating system reduces the amount of these materials needed, while improving their effectiveness.

In field tests conducted by AgZen, “we doubled the amount of product on kale and soybeans just by changing where the adjuvant was,” from mixed in to being a coating, Jayaprakash says. It’s convenient for farmers because “all they’re doing is changing their nozzle. They’re getting all their existing chemicals to work better, and they’re getting more product on the plant.”

And it’s not just for pesticides. “The really cool thing is this is useful for every chemistry that’s going on the leaf, be it an insecticide, a herbicide, a fungicide, or foliar nutrition,” Varanasi says. This year, they plan to introduce the new spray system on about 30,000 acres of cropland.

Varanasi says that with projected world population growth, “the amount of food production has got to double, and we are limited in so many resources, for example we cannot double the arable land. … This means that every acre we currently farm must become more efficient and able to do more with less.” These improved spraying technologies, for both monitoring the spraying and coating the droplets, Varanasi says, “I think is fundamentally changing agriculture.”

AgZen has recently raised $10 million in venture financing to support rapid commercial deployment of these technologies that can improve the control of chemical inputs into agriculture. “The knowledge we are gathering from every leaf, combined with our expertise in interfacial science and fluid mechanics, is giving us unparalleled insights into how chemicals are used and developed — and it’s clear that we can deliver value across the entire agrochemical supply chain,” Varanasi says  “Our mission is to use these technologies to deliver improved outcomes and reduced costs for the ag industry.” 

Early support for this research effort was provided by the Tata Center for Technology and Design, a part of the MIT Energy Initiative.

Two years ago, MIT professor of literature Arthur Bahr had one of the best days of his life. Sitting in the British Library, he was allowed to page through the Pearl-Manuscript, a singular bound volume from the 1300s containing the earliest versions of the masterly medieval poem “Pearl,” the famous tale “Sir Gawain and the Green Knight,” and two other poems.

Today, “Sir Gawain and the Green Knight” is commonly read in high school English classes. But it probably would have been lost to history without the survival of the Pearl-Manuscript, like the other works in the same volume. As it stands, no one knows who authored these texts. But one thing is clear: the surviving manuscript is a carefully crafted volume, with bespoke illustrations and the skilled use of parchment. This book is its own work of art.

“The Pearl-Manuscript is just as extraordinary and unusual and unexpected as the poems it contains,” Bahr says of the document, whose formal name is “British Library MS Cotton Nero A X/2.”

Bahr explores these ideas in a new book, “Chasing the Pearl-Manuscript: Speculation, Shapes, Delight,” published this month by the University of Chicago Press. In it, Bahr combines his deep knowledge of the volume’s texts with detailed examination of its physical qualities — thanks to technologies such as spectroscopy, which has revealed some manuscript secrets, as well as the good, old-fashioned scrutiny Bahr gave the book in person.

“My argument is that this physical object adds up to more than the sum of its parts, through its creative interplay of text, image, and materials,” Bahr says. “It is a coherent volume that evokes the concerns of the poems themselves. Most manuscripts are constructed in utilitarian ways, but not this one.”

Ode to the most beautiful poem

Bahr first encountered “Pearl” as an undergraduate at Amherst College, in a course taught by medievalist Howell D. Chickering. The poem is an intricate examination of Christian ethics; a father, whose daughter has died, dreams he is discussing the meaning of life with her.

“It is the most beautiful poem I have ever read,” Bahr says. “It blew me away, for its formal complexity, and for the really poignant human drama.” He adds: “It’s in some sense why I’m a medievalist.”

And since Bahr’s first book, “Fragments and Assemblages,” studies how medieval bound volumes were often collections of disparate documents, it was natural for him to apply this scholarly lens to the Pearl manuscript as well.

Most scholars think the Pearl manuscript has a single author — although we cannot be certain. After beginning with “Pearl,” the manuscript follows with two other poems, “Cleanness” and “Patience.” Closing the volume, “Sir Gawain and the Green Knight” is an eerie, surreal tale of courage and chivalry set in the (possibly fictional) court of King Arthur.

In the book, Bahr finds the four texts to be thematically linked, analyzing the “connective tissue” through which the “manuscript starts to cohere into a wrought, imperfect, temporally layered whole,” as he writes. Some of these links are broad, including recurring “challenges to our speculative faculties”; the works are full of seeming paradoxes and dreamscapes that test the reader’s interpretive capacity.

There are other ways the text seem aligned. “Pearl” and “Sir Gawain and the Green Knight” each have 101 stanzas. The texts have numerically consistent structures, in the case of “Pearl” based around the number 12. All but one of its stanzas has 12 lines (and Bahr suspects this imperfection is intentional, like a fine rug with a deliberate flaw, which may be the case for the “extra” 101st stanza). There are 36 lines per page. And from examining the manuscript in person, Bahr found 48 places with decorated initials, although we do not know whose.

“The more you look, the more you find,” Bahr says.

Materiality matters

Some of our knowledge about the Pearl-Manuscript is quite new: Spectroscopy has revealed that the volume originally had simple line drawings, which were later filled in with colored ink.

But there is no substitute for reading books in person. That took Bahr to London in 2023, where he was permitted an extended look at the Pearl-Manuscript in the flesh. Far from being a formality, that gave Bahr new insights.

For instance: The Pearl-Manuscript is written on parchment, which is animal skin. At a key point in the “Patience” poem, a reworking of the tale of Jonah and the whale, the parchment has been reversed, so that the “hair” side of the material faces up, rather than the “flesh” side; it is the only case of this in the manuscript.

“When you’re reading about Jonah being swallowed by the whale, you feel the hair follicles when you wouldn’t expect to,” Bahr says. “At precisely the moment when the poem is thematizing an unnatural reversal of inside and outside, you are feeling the other side of another animal.”

He adds: “The act of touching the Pearl-Manuscript really changed how I think this poem would have worked for the medieval reader.” In this vein, he says, “Materiality matters. Screens are enabling, and without the digital facsimile I could not have written this book, but they cannot ever replace the original. The ‘Patience’ chapter reinforces that.”

Ultimately, Bahr thinks the Pearl-Manuscript buttresses his view in the “Fragments and Assemblages” book, that the medieval reading experience was often bound up with the way volumes were physically constructed.

“My argument in ‘Fragments and Assemblages’ was that medieval readers and book constructors thought in a serious and often sophisticated way about how the material construction and the selection of the texts into a physical object made a difference — mattered — and had the potential to change the meanings of the texts,” he says.

Good grade on the group project

“Chasing the Pearl-Manuscript” has received praise from other scholars. Jessica Brantley, professor and chair of the English Department at Yale University, has said that Bahr “offers an adventurous multilayered reading of both text and book and provides an important reinterpretation of the codex and its poems.”

Daniel Wakelin of Oxford University has said that Bahr “sets out an authoritative reading of these poems” and presents “a bold model for studying material texts and literary works together.”

For his part, Bahr hopes to appeal to an array of readers, just as his courses on medieval literature appeal to students with an array of intellectual interests. In the making of his book, Bahr also credits two MIT students, Kelsey Glover and Madison Sneve, who helped the project through the Undergraduate Research Opportunities Program (UROP), studying the illustrations and distinctive manuscript markings, among other things.

“It’s a very MIT kind of poem in the sense that not only is the author, or authors, obsessed with math and geometry and numbers and proportion, they are also obsessed with artifact construction, with architectural details and physical craft,” Bahr says. “There’s a very ‘mens et manus’ quality to the poems that’s reflected in the manuscript,” he says, referring to MIT’s motto, “mind and hand.” “I think helps explain why these extraordinary MIT students helped me so much.”

If you filled out a March Madness bracket this month, you probably faced the same question with each college match-up: What gives one team an edge over another? Is it a team’s record through the regular season? Or the chemistry among its players? Maybe it’s the experience of its coaching staff or the buzz around a top scorer.

All of these factors play some role in a team’s chance to advance. But according to a new study by MIT researchers, there’s one member who consistently boosts their team’s performance: the data analyst.

The new study, which was published this month in the Journal of Sports Economics, quantifies the influence of basketball analytics investment on team performance. The study’s authors looked in particular at professional basketball and compared the  investment in data analytics on each NBA team with the team’s record of wins over 12 seasons. They found that indeed, teams that hired more analytics staff, and invested more in data analysis in general, tended to win more games.

Analytics department headcount had a positive and statistically significant effect on team wins even when accounting for other factors such as a team’s roster salary, the experience and chemistry among its players, the consistency of its coaching staff, and player injuries through each season. Even with all of these influences, the researchers found that the depth of a team’s data analytics bench, so to speak, was a consistent predictor of the team’s wins.

What’s more, they were able to quantify basketball analytics’ value, based on their impact on team wins. They found that for every four-fifths of one data analyst, a team gains one additional win in a season. Interestingly, a team can also gain one additional win by increasing its roster salary by $9.6 million. One way to read this is that one data analyst’s impact is worth at least $9 million.

“I don’t know of any analyst who’s being paid $9 million,” says study author Henry Wang, a graduate student in the MIT Sports Lab. “There is still a gap between how the player is being valued and how the analytics are being valued."

While the study focuses on professional basketball, the researchers say the findings are relevant beyond the NBA. They speculate that college teams that make use of data analytics may have an edge over those who don’t. (Take note, March Madness fans.) And the same likely goes for sports in general, along with any competitive field.

“This paper hits nicely not just in sports but beyond, with this question of: What is the tangible impact of big data analytics?” says co-author Arnab Sarker PhD ’25, a recent doctoral graduate of MIT’s Institute for Data, Systems and Society (IDSS). “Sports are a really nice, controlled place for analytics. But we’re also curious to what extent we can see these effects in general organizational performance.”

The study is also co-authored by Anette “Peko” Hosoi, the Pappalardo Professor of Mechanical Engineering at MIT.

Data return

Across the sports world, data analysts have grown in number and scope over the years. Sports analytics’ role in using data and stats to improve team performance was popularized in 2011 with the movie “Moneyball,” based on the 2003 book “Moneyball: The Art of Winning an Unfair Game” by Michael Lewis, who chronicled the 2002 Oakland Athletics and general manager Billy Beane’s use of baseball analytics to win games against wealthier Major League Baseball teams.

Since then, data analysis has expanded to many other sports, in an effort to make use of the varied and fast-paced sources of data, measurements, and statistics available today. In basketball, analysts can take on many roles, using data, for instance, to optimize a player’s health and minimize injury risk, and to predict a player’s performance to inform draft selection, free agency acquisition, and contract negotiations.

A data analyst’s work can also influence in-game strategy. Case in point: Over the last decade, NBA teams have strategically chosen to shift to shooting longer-range three-pointers, since Philadelphia 76ers President of Basketball Operations Daryl Morey SM ’00 determined that statistically, shooting more three-pointers wins more games. Today, each of the 30 NBA teams employs at least one basketball analytics staffer. And yet, while a data analyst’s job is entirely based on data, there is not much data on the impact of analysts themselves.

“Teams and leagues are spending millions of dollars on embracing analytical tools without a real sense of return-on-investment,” Wang notes.

Numbers value

The MIT researchers aimed in their new study to quantify the influence of NBA team analysts, specifically on winning games. To do so, they looked to major sources of sports data such as ESPN.com, and NBAstuffer.com, a website that hosts a huge amount of stats on NBA games and team stats, including hired basketball analytics staff, that the website’s managers compile based on publicly available data, such as from official team press releases and staff directories, as well as LinkedIn and X profiles, and news and industry reports.

For their new study, Wang and his colleagues gathered data on each of the 30 NBA teams, over a period from 2009 to 2023, 2009 being the year that NBAstuffer.com started compiling team data. For every team in each season during this period, the researchers recorded an “analyst headcount,” meaning the number of basketball operations analytics staff employed by a team. They considered an analyst to be data analysts, software engineers, sports scientists, directors of research, and other technical positions by title, but also staff members who aren’t formally analysts but may be known to be particularly active in the basketball analytics community. In general, they found that in 2009, a total of 10 data analysts were working across the NBA. In 2023, that number ballooned to 132, with some teams employing more analysts than others.

“What we’re trying to measure is a team’s level of investment in basketball analytics,” Wang explains. “The best measure would be if every team told us exactly how much money they spent every year on their R&D and data infrastructure and analysts. But they’re not going to do that. So headcount is the next best thing.”

In addition to analytics headcount, the researchers also compiled data on other win-influencing variables, such as roster salary (Does a higher-paid team win more games?), roster experience (Does a team with more veterans win more games?), consistent coaching (Did a new coach shake up a team’s win record?) and season injuries (How did a team’s injuries affect its wins?). The researchers also noted “road back-to-backs,” or the number of times a team had to play consecutive away games (Does the wear and tear of constant travel impact wins?).

The researchers plugged all this data into a “two-way fixed effects” model to estimate the relative effect that each variable has on the number of additional games a team can win in a season.

“The model learns all these effects, so we can see, for instance, the tradeoff between analyst and roster salary when contributing to win total,” Wang explains.

Their finding that teams with a higher analytics headcount tended to win more games wasn’t entirely surprising.

“We’re still at a point where the analyst is undervalued,” Wang says. “There probably is a sweet spot, in terms of headcount and wins. You can’t hire 100 analysts and expect to go in 82-and-0 next season. But right now a lot of teams are still below that sweet spot, and this competitive advantage that analytics offers has yet to be fully harvested.”

Next time you cross a crowded plaza, crosswalk, or airport concourse, take note of the pedestrian flow. Are people walking in orderly lanes, single-file, to their respective destinations? Or is it a haphazard tangle of personal trajectories, as people dodge and weave through the crowd?

MIT instructor Karol Bacik and his colleagues studied the flow of human crowds and developed a first-of-its-kind way to predict when pedestrian paths will transition from orderly to entangled. Their findings may help inform the design of public spaces that promote safe and efficient thoroughfares.

In a paper appearing this week in the Proceedings of the National Academy of Sciences, the researchers consider a common scenario in which pedestrians navigate a busy crosswalk. The team analyzed the scenario through mathematical analysis and simulations, considering the many angles at which individuals may cross and the dodging maneuvers they may make as they attempt to reach their destinations while avoiding bumping into other pedestrians along the way.

The researchers also carried out controlled crowd experiments and studied how real participants walked through a crowd to reach certain locations. Through their mathematical and experimental work, the team identified a key measure that determines whether pedestrian traffic is ordered, such that clear lanes form in the flow, or disordered, in which there are no discernible paths through the crowd. Called “angular spread,” this parameter describes the number of people walking in different directions.

If a crowd has a relatively small angular spread, this means that most pedestrians walk in opposite directions and meet the oncoming traffic head-on, such as in a crosswalk. In this case, more orderly, lane-like traffic is likely. If, however, a crowd has a larger angular spread, such as in a concourse, it means there are many more directions that pedestrians can take to cross, with more chance for disorder.

In fact, the researchers calculated the point at which a moving crowd can transition from order to disorder. That point, they found, was an angular spread of around 13 degrees, meaning that if pedestrians don’t walk straight across, but instead an average pedestrian veers off at an angle larger than 13 degrees, this can tip a crowd into disordered flow.

“This all is very commonsense,” says Bacik, who is a instructor of applied mathematics at MIT. “The question is whether we can tackle it precisely and mathematically, and where the transition is. Now we have a way to quantify when to expect lanes — this spontaneous, organized, safe flow — versus disordered, less efficient, potentially more dangerous flow.”

The study’s co-authors include Grzegorz Sobota and Bogdan Bacik of the Academy of Physical Education in Katowice, Poland, and Tim Rogers at the University of Bath in the United Kingdom.

Right, left, center

Bacik, who is trained in fluid dynamics and granular flow, came to study pedestrian flow during 2021, when he and his collaborators looked into the impacts of social distancing, and ways in which people might walk among each other while maintaining safe distances. That work inspired them to look more generally into the dynamics of crowd flow.

In 2023, he and his collaborators explored “lane formation,” a phenomenon by which particles, grains, and, yes, people have been observed to spontaneously form lanes, moving in single-file when forced to cross a region from two opposite directions. In that work, the team identified the mechanism by which such lanes form, which Bacik sums up as “an imbalance of turning left versus right.” Essentially, they found that as soon as something in a crowd starts to look like a lane, individuals around that fledgling lane either join up, or are forced to either side of it, walking parallel to the original lane, which others can follow. In this way, a crowd can spontaneously organize into regular, structured lanes.

“Now we’re asking, how robust is this mechanism?” Bacik says. “Does it only work in this very idealized situation, or can lane formation tolerate some imperfections, such as some people not going perfectly straight, as they might do in a crowd?”

Lane change

For their new study, the team looked to identify a key transition in crowd flow: When do pedestrians switch from orderly, lane-like traffic, to less organized, messy flow? The researchers first probed the question mathematically, with an equation that is typically used to describe fluid flow, in terms of the average motion of many individual molecules.

“If you think about the whole crowd flowing, rather than individuals, you can use fluid-like descriptions,” Bacik explains. “It’s this art of averaging, where, even if some people may cross more assertively than others, these effects are likely to average out in a sufficiently large crowd. If you only care about the global characteristics like, are there lanes or not, then you can make predictions without detailed knowledge of everyone in the crowd.”

Bacik and his colleagues used equations of fluid flow, and applied them to the scenario of pedestrians flowing across a crosswalk. The team tweaked certain parameters in the equation, such as the width of the fluid channel (in this case, the crosswalk), and the angle at which molecules (or people) flowed across, along with various directions that people can “dodge,” or move around each other to avoid colliding.

Based on these calculations, the researchers found that pedestrians in a crosswalk are more likely to form lanes, when they walk relatively straight across, from opposite directions. This order largely holds until people start veering across at more extreme angles. Then, the equation predicts that the pedestrian flow is likely to be disordered, with few to no lanes forming.

The researchers were curious to see whether the math bears out in reality. For this, they carried out experiments in a gymnasium, where they recorded the movements of pedestrians using an overhead camera. Each volunteer wore a paper hat, depicting a unique barcode that the overhead camera could track.

In their experiments, the team assigned volunteers various start and end positions along opposite sides of a simulated crosswalk, and tasked them with simultaneously walking across the crosswalk to their target location without bumping into anyone. They repeated the experiment many times, each time having volunteers assume different start and end positions. In the end, the researchers were able to gather visual data of multiple crowd flows, with pedestrians taking many different crossing angles.

When they analyzed the data and noted when lanes spontaneously formed, and when they did not, the team found that, much like the equation predicted, the angular spread mattered. Their experiments confirmed that the transition from ordered to disordered flow occurred somewhere around the theoretically predicted 13 degrees. That is, if an average person veered more than 13 degrees away from straight ahead, the pedestrian flow could tip into disorder, with little lane formation. What’s more, they found that the more disorder there is in a crowd, the less efficiently it moves.

The team plans to test their predictions on real-world crowds and pedestrian thoroughfares.

“We would like to analyze footage and compare that with our theory,” Bacik says. “And we can imagine that, for anyone designing a public space, if they want to have a safe and efficient pedestrian flow, our work could provide a simpler guideline, or some rules of thumb.”

This work is supported, in part, by the Engineering and Physical Sciences Research Council of UK Research and Innovation.

Life takes shape with the motion of a single cell. In response to signals from certain proteins and enzymes, a cell can start to move and shake, leading to contractions that cause it to squeeze, pinch, and eventually divide. As daughter cells follow suit down the generational line, they grow, differentiate, and ultimately arrange themselves into a fully formed organism.

Now MIT scientists have used light to control how a single cell jiggles and moves during its earliest stage of development. The team studied the motion of egg cells produced by starfish — an organism that scientists have long used as a classic model for understanding cell growth and development.

The researchers focused on a key enzyme that triggers a cascade of motion within a starfish egg cell. They genetically designed a light-sensitive version of the same enzyme, which they injected into egg cells, and then stimulated the cells with different patterns of light.

They found that the light successfully triggered the enzyme, which in turn prompted the cells to jiggle and move in predictable patterns. For instance, the scientists could stimulate cells to exhibit small pinches or sweeping contractions, depending on the pattern of light they induced. They could even shine light at specific points around a cell to stretch its shape from a circle to a square.

Their results, appearing today in the journal Nature Physics, provide scientists with a new optical tool for controlling cell shape in its earliest developmental stages. Such a tool, they envision, could guide the design of synthetic cells, such as therapeutic “patch” cells that contract in response to light signals to help close wounds, or drug-delivering “carrier” cells that release their contents only when illuminated at specific locations in the body. Overall, the researchers see their findings as a new way to probe how life takes shape from a single cell.

“By revealing how a light-activated switch can reshape cells in real time, we’re uncovering basic design principles for how living systems self-organize and evolve shape,” says the study’s senior author, Nikta Fakhri, associate professor of physics at MIT. “The power of these tools is that they are guiding us to decode all these processes of growth and development, to help us understand how nature does it.”

The study’s MIT authors include first author Jinghui Liu, Yu-Chen Chao, and Tzer Han Tan; along with Tom Burkart, Alexander Ziepke, and Erwin Frey of Ludwig Maximilian University of Munich; John Reinhard of Saarland University; and S. Zachary Swartz of the Whitehead Institute for Biomedical Research.

Cell circuitry

Fakhri’s group at MIT studies the physical dynamics that drive cell growth and development. She is particularly interested in symmetry, and the processes that govern how cells follow or break symmetry as they grow and divide. The five-limbed starfish, she says, is an ideal organism for exploring such questions of growth, symmetry, and early development.

“A starfish is a fascinating system because it starts with a symmetrical cell and becomes a bilaterally symmetric larvae at early stages, and then develops into pentameral adult symmetry,” Fakhri says. “So there’s all these signaling processes that happen along the way to tell the cell how it needs to organize.”

Scientists have long studied the starfish and its various stages of development. Among many revelations, researchers have discovered a key “circuitry” within a starfish egg cell that controls its motion and shape. This circuitry involves an enzyme, GEF, that naturally circulates in a cell’s cytoplasm. When this enzyme is activated, it induces a change in a protein, called Rho, that is known to be essential for regulating cell mechanics.

When the GEF enzyme stimulates Rho, it causes the protein to switch from an essentially free-floating state to a state that binds the protein to the cell’s membrane. In this membrane-bound state, the protein then triggers the growth of microscopic, muscle-like fibers that thread out across the membrane and subsequently twitch, enabling the cell to contract and move. 

In previous work, Fakhri’s group showed that a cell’s movements can be manipulated by varying the cell’s concentrations of GEF enzyme: The more enzyme they introduced into a cell, the more contractions the cell would exhibit.

“This whole idea made us think whether it’s possible to hack this circuitry, to not just change a cell’s pattern of movements but get a desired mechanical response,” Fakhri says.

Lights and action

To precisely manipulate a cell’s movements, the team looked to optogenetics — an approach that involves genetically engineering cells and cellular components such as proteins and enzymes, such that they activate in response to light.

Using established optogenetic techniques, the researchers developed a light-sensitive version of the GEF enzyme. From this engineered enzyme, they isolated its mRNA — essentially, the genetic blueprint for building the enzyme. They then injected this blueprint into egg cells that the team harvested from a single starfish ovary, which can hold millions of unfertilized cells. The cells, infused with the new mRNA, then began to produce light-sensitive GEF enzymes on their own.

In experiments, the researchers then placed each enzyme-infused egg cell under a microscope and shone light onto the cell in different patterns and from different points along the cell’s periphery. They took videos of the cell’s movements in response.

They found that when they aimed the light in specific points, the GEF enzyme became activated and recruited Rho protein to the light-targeted sites. There, the protein then set off its characteristic cascade of muscle-like fibers that pulled or pinched the cell in the same, light-stimulated spots. Much like pulling the strings of a marionette, they were able to control the cell’s movements, for instance directing it to morph into various shapes, including a square.

Surprisingly, they also found they could stimulate the cell to undergo sweeping contractions by shining a light in a single spot, exceeding a certain threshold of enzyme concentration.

“We realized this Rho-GEF circuitry is an excitable system, where a small, well-timed stimulus can trigger a large, all-or-nothing response,” Fakhri says. “So we can either illuminate the whole cell, or just a tiny place on the cell, such that enough enzyme is recruited to that region so the system gets kickstarted to contract or pinch on its own.”

The researchers compiled their observations and derived a theoretical framework to predict how a cell’s shape will change, given how it is stimulated with light. The framework, Fakhri says, opens a window into “the ‘excitability’ at the heart of cellular remodeling, which is a fundamental process in embryo development and wound healing.”

She adds: “This work provides a blueprint for designing ‘programmable’ synthetic cells, letting researchers orchestrate shape changes at will for future biomedical applications.”

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

MIT engineers have devised a new way to deliver certain drugs in higher doses with less pain, by injecting them as a suspension of tiny crystals. Once under the skin, the crystals assemble into a drug “depot” that could last for months or years, eliminating the need for frequent drug injections.

This approach could prove useful for delivering long-lasting contraceptives or other drugs that need to be given for extended periods of time. Because the drugs are dispersed in a suspension before injection, they can be administered through a narrow needle that is easier for patients to tolerate.

“We showed that we can have very controlled, sustained delivery, likely for multiple months and even years through a small needle,” says Giovanni Traverso, an associate professor of mechanical engineering at MIT, a gastroenterologist at Brigham and Women’s Hospital (BWH), an associate member of the Broad Institute, and the senior author of the study.

The lead authors of the paper, which appears today in Nature Chemical Engineering, are former MIT and BWH postdoc Vivian Feig, who is now an assistant professor of mechanical engineering at Stanford University; MIT graduate student Sanghyun Park; and Pier Rivano, a former visiting research scholar in Traverso’s lab.

Easier injections

This project began as part of an effort funded by the Gates Foundation to expand contraceptive options, particularly in developing nations.

“The overarching goal is to give women access to a lot of different formats for contraception that are easy to administer, compatible with being used in the developing world, and have a range of different timeframes of durations of action,” Feig says. “In our particular project, we were interested in trying to combine the benefits of long-acting implants with the ease of self-administrable injectables.”

There are marketed injectable suspensions available in the United States and other countries, but these drugs are dispersed throughout the tissue after injection, so they only work for about three months. Other injectable products have been developed that can form longer-lasting depots under the skin, but these typically require the addition of precipitating polymers that can make up 23 to 98 percent of the solution by weight, which can make the drug more difficult to inject.

The MIT and BWH team wanted to create a formulation that could be injected through a small-gauge needle and last for at least six months and up to two years. They began working with a contraceptive drug called levonorgestrel, a hydrophobic molecule that can form crystals. The team discovered that suspending these crystals in a particular organic solvent caused the crystals to assemble into a highly compact implant after injection. Because this depot could form without needing large amounts of polymer, the drug formulation could still be easily injected through a narrow-gauge needle.

The solvent, benzyl benzoate, is biocompatible and has been previously used as an additive to injectable drugs. The team found that the solvent’s poor ability to mix with biological fluids is what allows the solid drug crystals to self-assemble into a depot under the skin after injection.

“The solvent is critical because it allows you to inject the fluid through a small needle, but once in place, the crystals self-assemble into a drug depot,” Traverso says.

By altering the density of the depot, the researchers can tune the rate at which the drug molecules are released into the body. In this study, the researchers showed they could change the density by adding small amounts of a polymer such as polycaprolactone, a biodegradable polyester.

“By incorporating a very small amount of polymers — less than 1.6 percent by weight — we can modulate the drug release rate, extending its duration while maintaining injectability. This demonstrates the tunability of our system, which can be engineered to accommodate a broader range of contraceptive needs as well as tailored dosing regimens for other therapeutic applications,” Park says.

Stable drug depots

The researchers tested their approach by injecting the drug solution subcutaneously in rats and showed that the drug depots could remain stable and release drug gradually for three months. After the three-month study ended, about 85 percent of the drug remained in the depots, suggesting that they could continue releasing the drugs for a much longer period of time.

“We anticipate that the depots could last for more than a year, based on our post-analysis of preclinical data. Follow-up studies are underway to further validate their efficacy beyond this initial proof-of-concept,” Park says.

Once the drug depots form, they are compact enough to be retrievable, allowing for surgical removal if treatment needs to be halted before the drug is fully released.

This approach could also lend itself to delivering drugs to treat neuropsychiatric conditions as well as HIV and tuberculosis, the researchers say. They are now moving toward assessing its translation to humans by conducting advanced preclinical studies to evaluate self-assembly in a more clinically relevant skin environment. “This is a very simple system in that it’s basically a solvent, the drug, and then you can add a little bit of bioresorbable polymer. Now we’re considering which indications do we go after: Is it contraception? Is it others? These are some of the things that we’re starting to look into as part of the next steps toward translation to humans,” Traverso says.

The research was funded, in part, by the Gates Foundation, the Karl van Tassel Career Development Professorship, the MIT Department of Mechanical Engineering, a Schmidt Science Fellows postdoctoral fellowship, the Rhodes Trust, a Takeda Fellowship, a Warren M. Rohsenow Fellowship, and a Kwangjeong Educational Foundation Fellowship.

Quantum computers have the potential to solve complex problems that would be impossible for the most powerful classical supercomputer to crack.

Just like a classical computer has separate, yet interconnected, components that must work together, such as a memory chip and a CPU on a motherboard, a quantum computer will need to communicate quantum information between multiple processors.

Current architectures used to interconnect superconducting quantum processors are “point-to-point” in connectivity, meaning they require a series of transfers between network nodes, with compounding error rates.

On the way to overcoming these challenges, MIT researchers developed a new interconnect device that can support scalable, “all-to-all” communication, such that all superconducting quantum processors in a network can communication directly with each other.

They created a network of two quantum processors and used their interconnect to send microwave photons back and forth on demand in a user-defined direction. Photons are particles of light that can carry quantum information.

The device includes a superconducting wire, or waveguide, that shuttles photons between processors and can be routed as far as needed. The researchers can couple any number of modules to it, efficiently transmitting information between a scalable network of processors.

They used this interconnect to demonstrate remote entanglement, a type of correlation between quantum processors that are not physically connected. Remote entanglement is a key step toward developing a powerful, distributed network of many quantum processors.

“In the future, a quantum computer will probably need both local and nonlocal interconnects. Local interconnects are natural in arrays of superconducting qubits. Ours allows for more nonlocal connections. We can send photons at different frequencies, times, and in two propagation directions, which gives our network more flexibility and throughput,” says Aziza Almanakly, an electrical engineering and computer science graduate student in the Engineering Quantum Systems group of the Research Laboratory of Electronics (RLE) and lead author of a paper on the interconnect.

Her co-authors include Beatriz Yankelevich, a graduate student in the EQuS Group; senior author William D. Oliver, the Henry Ellis Warren (1894) Professor of Electrical Engineering and Computer Science (EECS) and professor of Physics, director of the Center for Quantum Engineering, and associate director of RLE; and others at MIT and Lincoln Laboratory. The research appears today in Nature Physics.

A scalable architecture

The researchers previously developed a quantum computing module, which enabled them to send information-carrying microwave photons in either direction along a waveguide.

In the new work, they took that architecture a step further by connecting two modules to a waveguide in order to emit photons in a desired direction and then absorb them at the other end.

Each module is composed of four qubits, which serve as an interface between the waveguide carrying the photons and the larger quantum processors.

The qubits coupled to the waveguide emit and absorb photons, which are then transferred to nearby data qubits.

The researchers use a series of microwave pulses to add energy to a qubit, which then emits a photon. Carefully controlling the phase of those pulses enables a quantum interference effect that allows them to emit the photon in either direction along the waveguide. Reversing the pulses in time enables a qubit in another module any arbitrary distance away to absorb the photon.

“Pitching and catching photons enables us to create a ‘quantum interconnect’ between nonlocal quantum processors, and with quantum interconnects comes remote entanglement,” explains Oliver.

“Generating remote entanglement is a crucial step toward building a large-scale quantum processor from smaller-scale modules. Even after that photon is gone, we have a correlation between two distant, or ‘nonlocal,’ qubits. Remote entanglement allows us to take advantage of these correlations and perform parallel operations between two qubits, even though they are no longer connected and may be far apart,” Yankelevich explains.

However, transferring a photon between two modules is not enough to generate remote entanglement. The researchers need to prepare the qubits and the photon so the modules “share” the photon at the end of the protocol.

Generating entanglement

The team did this by halting the photon emission pulses halfway through their duration. In quantum mechanical terms, the photon is both retained and emitted. Classically, one can think that half-a-photon is retained and half is emitted.

Once the receiver module absorbs that “half-photon,” the two modules become entangled.

But as the photon travels, joints, wire bonds, and connections in the waveguide distort the photon and limit the absorption efficiency of the receiving module.

To generate remote entanglement with high enough fidelity, or accuracy, the researchers needed to maximize how often the photon is absorbed at the other end.

“The challenge in this work was shaping the photon appropriately so we could maximize the absorption efficiency,” Almanakly says.

They used a reinforcement learning algorithm to “predistort” the photon. The algorithm optimized the protocol pulses in order to shape the photon for maximal absorption efficiency.

When they implemented this optimized absorption protocol, they were able to show photon absorption efficiency greater than 60 percent.

This absorption efficiency is high enough to prove that the resulting state at the end of the protocol is entangled, a major milestone in this demonstration.

“We can use this architecture to create a network with all-to-all connectivity. This means we can have multiple modules, all along the same bus, and we can create remote entanglement among any pair of our choosing,” Yankelevich says.

In the future, they could improve the absorption efficiency by optimizing the path over which the photons propagate, perhaps by integrating modules in 3D instead of having a superconducting wire connecting separate microwave packages. They could also make the protocol faster so there are fewer chances for errors to accumulate.

“In principle, our remote entanglement generation protocol can also be expanded to other kinds of quantum computers and bigger quantum internet systems,” Almanakly says.

This work was funded, in part, by the U.S. Army Research Office, the AWS Center for Quantum Computing, and the U.S. Air Force Office of Scientific Research. 

The ability to generate high-quality images quickly is crucial for producing realistic simulated environments that can be used to train self-driving cars to avoid unpredictable hazards, making them safer on real streets.

But the generative artificial intelligence techniques increasingly being used to produce such images have drawbacks. One popular type of model, called a diffusion model, can create stunningly realistic images but is too slow and computationally intensive for many applications. On the other hand, the autoregressive models that power LLMs like ChatGPT are much faster, but they produce poorer-quality images that are often riddled with errors.

Researchers from MIT and NVIDIA developed a new approach that brings together the best of both methods. Their hybrid image-generation tool uses an autoregressive model to quickly capture the big picture and then a small diffusion model to refine the details of the image.

Their tool, known as HART (short for hybrid autoregressive transformer), can generate images that match or exceed the quality of state-of-the-art diffusion models, but do so about nine times faster.

The generation process consumes fewer computational resources than typical diffusion models, enabling HART to run locally on a commercial laptop or smartphone. A user only needs to enter one natural language prompt into the HART interface to generate an image.

HART could have a wide range of applications, such as helping researchers train robots to complete complex real-world tasks and aiding designers in producing striking scenes for video games.

“If you are painting a landscape, and you just paint the entire canvas once, it might not look very good. But if you paint the big picture and then refine the image with smaller brush strokes, your painting could look a lot better. That is the basic idea with HART,” says Haotian Tang SM ’22, PhD ’25, co-lead author of a new paper on HART.

He is joined by co-lead author Yecheng Wu, an undergraduate student at Tsinghua University; senior author Song Han, an associate professor in the MIT Department of Electrical Engineering and Computer Science (EECS), a member of the MIT-IBM Watson AI Lab, and a distinguished scientist of NVIDIA; as well as others at MIT, Tsinghua University, and NVIDIA. The research will be presented at the International Conference on Learning Representations.

The best of both worlds

Popular diffusion models, such as Stable Diffusion and DALL-E, are known to produce highly detailed images. These models generate images through an iterative process where they predict some amount of random noise on each pixel, subtract the noise, then repeat the process of predicting and “de-noising” multiple times until they generate a new image that is completely free of noise.

Because the diffusion model de-noises all pixels in an image at each step, and there may be 30 or more steps, the process is slow and computationally expensive. But because the model has multiple chances to correct details it got wrong, the images are high-quality.

Autoregressive models, commonly used for predicting text, can generate images by predicting patches of an image sequentially, a few pixels at a time. They can’t go back and correct their mistakes, but the sequential prediction process is much faster than diffusion.

These models use representations known as tokens to make predictions. An autoregressive model utilizes an autoencoder to compress raw image pixels into discrete tokens as well as reconstruct the image from predicted tokens. While this boosts the model’s speed, the information loss that occurs during compression causes errors when the model generates a new image.

With HART, the researchers developed a hybrid approach that uses an autoregressive model to predict compressed, discrete image tokens, then a small diffusion model to predict residual tokens. Residual tokens compensate for the model’s information loss by capturing details left out by discrete tokens.

“We can achieve a huge boost in terms of reconstruction quality. Our residual tokens learn high-frequency details, like edges of an object, or a person’s hair, eyes, or mouth. These are places where discrete tokens can make mistakes,” says Tang.

Because the diffusion model only predicts the remaining details after the autoregressive model has done its job, it can accomplish the task in eight steps, instead of the usual 30 or more a standard diffusion model requires to generate an entire image. This minimal overhead of the additional diffusion model allows HART to retain the speed advantage of the autoregressive model while significantly enhancing its ability to generate intricate image details.

“The diffusion model has an easier job to do, which leads to more efficiency,” he adds.

Outperforming larger models

During the development of HART, the researchers encountered challenges in effectively integrating the diffusion model to enhance the autoregressive model. They found that incorporating the diffusion model in the early stages of the autoregressive process resulted in an accumulation of errors. Instead, their final design of applying the diffusion model to predict only residual tokens as the final step significantly improved generation quality.

Their method, which uses a combination of an autoregressive transformer model with 700 million parameters and a lightweight diffusion model with 37 million parameters, can generate images of the same quality as those created by a diffusion model with 2 billion parameters, but it does so about nine times faster. It uses about 31 percent less computation than state-of-the-art models.

Moreover, because HART uses an autoregressive model to do the bulk of the work — the same type of model that powers LLMs — it is more compatible for integration with the new class of unified vision-language generative models. In the future, one could interact with a unified vision-language generative model, perhaps by asking it to show the intermediate steps required to assemble a piece of furniture.

“LLMs are a good interface for all sorts of models, like multimodal models and models that can reason. This is a way to push the intelligence to a new frontier. An efficient image-generation model would unlock a lot of possibilities,” he says.

In the future, the researchers want to go down this path and build vision-language models on top of the HART architecture. Since HART is scalable and generalizable to multiple modalities, they also want to apply it for video generation and audio prediction tasks.

This research was funded, in part, by the MIT-IBM Watson AI Lab, the MIT and Amazon Science Hub, the MIT AI Hardware Program, and the U.S. National Science Foundation. The GPU infrastructure for training this model was donated by NVIDIA. 

It’s difficult to build devices that replicate the fluid, precise motion of humans, but that might change if we could pull a few (literal) strings.

At least, that’s the idea behind “cable-driven” mechanisms in which running a string through an object generates streamlined movement across an object’s different parts. Take a robotic finger, for example: You could embed a cable through the palm to the fingertip of this object and then pull it to create a curling motion.

While cable-driven mechanisms can create real-time motion to make an object bend, twist, or fold, they can be complicated and time-consuming to assemble by hand. To automate the process, researchers from MIT’s Computer Science and Artificial Intelligence Laboratory (CSAIL) have developed an all-in-one 3D printing approach called “Xstrings.” Part design tool, part fabrication method, Xstrings can embed all the pieces together and produce a cable-driven device, saving time when assembling bionic robots, creating art installations, or working on dynamic fashion designs.

In a paper to be presented at the 2025 Conference on Human Factors in Computing Systems (CHI2025), the researchers used Xstrings to print a range of colorful and unique objects that included a red walking lizard robot, a purple wall sculpture that can open and close like a peacock’s tail, a white tentacle that curls around items, and a white claw that can ball up into a fist to grab objects.

To fabricate these eye-catching mechanisms, Xstrings allows users to fully customize their designs in a software program, sending them to a multi-material 3D printer to bring that creation to life. You can automatically print all the device’s parts in their desired locations in one step, including the cables running through it and the joints that enable its intended motion.

MIT CSAIL postdoc and lead author Jiaji Li says that Xstrings can save engineers time and energy, reducing 40 percent of total production time compared to doing things manually. “Our innovative method can help anyone design and fabricate cable-driven products with a desktop bi-material 3D printer,” says Li.

A new twist on cable-driven fabrication

To use the Xstrings program, users first input a design with specific dimensions, like a rectangular cube divided into smaller pieces with a hole in the middle of each one. You can then choose which way its parts move by selecting different “primitives:” bending, coiling (like a spring), twisting (like a screw), or compressing — and the angle of these motions.

For even more elaborate creations, users can incorporate multiple primitives to create intriguing combinations of motions. If you wanted to make a toy snake, you could include several twists to create a “series” combo, in which a single cord drives a sequence of motions. To create the robot claw, the team embedded multiple cables into a “parallel” combination, where several strings are embedded, to enable each finger to close up into a fist.

Beyond fine-tuning the way cable-driven mechanisms move, Xstrings also facilitates how cables are integrated into the object. Users can choose exactly how the strings are secured, in terms of where the “anchor” (endpoint), “threaded areas” (or holes within the structure that the cord passes through), and “exposed point” (where you’d pull to operate the device) are located. With a robot finger, for instance, you could choose the anchor to be located at the fingertip, with a cable running through the finger and a pull tag exposed at the other end.

Xstrings also supports diverse joint designs by automatically placing components that are elastic, compliant, or mechanical. This allows the cable to turn as needed as it completes the device’s intended motion.

Driving unique designs across robotics, art, and beyond

Once users have simulated their digital blueprint for a cable-driven item, they can bring it to life via fabrication. Xstrings can send your design to a fused deposition modeling 3D printer, where plastic is melted down into a nozzle before the filaments are poured out to build structures up layer by layer.

Xstrings uses this technique to lay out cables horizontally and build around them. To ensure their method would successfully print cable-driven mechanisms, the researchers carefully tested their materials and printing conditions.

For example, the researchers found that their strings only broke after being pulled up and down by a mechanical device more than 60,000 times. In another test, the team discovered that printing at 260 degrees Celsius with a speed of 10-20 millimeters per second was ideal for producing their many creative items.

“The Xstrings software can bring a variety of ideas to life,” says Li. “It enables you to produce a bionic robot device like a human hand, mimicking our own gripping capabilities. You can also create interactive art pieces, like a cable-driven sculpture with unique geometries, and clothes with adjustable flaps. One day, this technology could enable the rapid, one-step creation of cable-driven robots in outer space, even within highly confined environments such as space stations or extraterrestrial bases.”

The team’s approach offers plenty of flexibility and a noticeable speed boost to fabricating cable-driven objects. It creates objects that are rigid on the outside, but soft and flexible on the inside; in the future, they may look to develop objects that are soft externally but rigid internally, much like humans’ skin and bones. They’re also considering using more resilient cables, and, instead of just printing strings horizontally, embedding ones that are angled or even vertical.

Li wrote the paper with Zhejiang University master’s student Shuyue Feng; Tsinghua University master’s student Yujia Liu; Zhejiang University assistant professor and former MIT Media Lab visiting researcher Guanyun Wang; and three CSAIL members: Maxine Perroni-Scharf, an MIT PhD student in electrical engineering and computer science; Emily Guan, a visiting researcher; and senior 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.

This research was supported, in part, by a postdoctoral research fellowship from Zhejiang University, and the MIT-GIST Program.

Within the human brain, a network of regions has evolved to process language. These regions are consistently activated whenever people listen to their native language or any language in which they are proficient.

A new study by MIT researchers finds that this network also responds to languages that are completely invented, such as Esperanto, which was created in the late 1800s as a way to promote international communication, and even to languages made up for television shows such as “Star Trek” and “Game of Thrones.”

To study how the brain responds to these artificial languages, MIT neuroscientists convened nearly 50 speakers of these languages over a single weekend. Using functional magnetic resonance imaging (fMRI), the researchers found that when participants listened to a constructed language in which they were proficient, the same brain regions lit up as those activated when they processed their native language.

“We find that constructed languages very much recruit the same system as natural languages, which suggests that the key feature that is necessary to engage the system may have to do with the kinds of meanings that both kinds of languages can express,” says Evelina Fedorenko, an associate professor of neuroscience at MIT, a member of MIT’s McGovern Institute for Brain Research and the senior author of the study.

The findings help to define some of the key properties of language, the researchers say, and suggest that it’s not necessary for languages to have naturally evolved over a long period of time or to have a large number of speakers.

“It helps us narrow down this question of what a language is, and do it empirically, by testing how our brain responds to stimuli that might or might not be language-like,” says Saima Malik-Moraleda, an MIT postdoc and the lead author of the paper, which appears this week in the Proceedings of the National Academy of Sciences.

Convening the conlang community

Unlike natural languages, which evolve within communities and are shaped over time, constructed languages, or “conlangs,” are typically created by one person who decides what sounds will be used, how to label different concepts, and what the grammatical rules are.

Esperanto, the most widely spoken conlang, was created in 1887 by L.L. Zamenhof, who intended it to be used as a universal language for international communication. Currently, it is estimated that around 60,000 people worldwide are proficient in Esperanto.

In previous work, Fedorenko and her students have found that computer programming languages, such as Python — another type of invented language — do not activate the brain network that is used to process natural language. Instead, people who read computer code rely on the so-called multiple demand network, a brain system that is often recruited for difficult cognitive tasks.

Fedorenko and others have also investigated how the brain responds to other stimuli that share features with language, including music and nonverbal communication such as gestures and facial expressions.

“We spent a lot of time looking at all these various kinds of stimuli, finding again and again that none of them engage the language-processing mechanisms,” Fedorenko says. “So then the question becomes, what is it that natural languages have that none of those other systems do?”

That led the researchers to wonder if artificial languages like Esperanto would be processed more like programming languages or more like natural languages. Similar to programming languages, constructed languages are created by an individual for a specific purpose, without natural evolution within a community. However, unlike programming languages, both conlangs and natural languages can be used to convey meanings about the state of the external world or the speaker’s internal state.

To explore how the brain processes conlangs, the researchers invited speakers of Esperanto and several other constructed languages to MIT for a weekend conference in November 2022. The other languages included Klingon (from “Star Trek”), Na’vi (from “Avatar”), and two languages from “Game of Thrones” (High Valyrian and Dothraki). For all of these languages, there are texts available for people who want to learn the language, and for Esperanto, Klingon, and High Valyrian, there is even a Duolingo app available.

“It was a really fun event where all the communities came to participate, and over a weekend, we collected all the data,” says Malik-Moraleda, who co-led the data collection effort with former MIT postbac Maya Taliaferro, now a PhD student at New York University.

During that event, which also featured talks from several of the conlang creators, the researchers used fMRI to scan 44 conlang speakers as they listened to sentences from the constructed language in which they were proficient. The creators of these languages — who are co-authors on the paper — helped construct the sentences that were presented to the participants.

While in the scanner, the participants also either listened to or read sentences in their native language, and performed some nonlinguistic tasks for comparison. The researchers found that when people listened to a conlang, the same language regions in the brain were activated as when they listened to their native language.

Common features

The findings help to identify some of the key features that are necessary to recruit the brain’s language processing areas, the researchers say. One of the main characteristics driving language responses seems to be the ability to convey meanings about the interior and exterior world — a trait that is shared by natural and constructed languages, but not programming languages.

“All of the languages, both natural and constructed, express meanings related to inner and outer worlds. They refer to objects in the world, to properties of objects, to events,” Fedorenko says. “Whereas programming languages are much more similar to math. A programming language is a symbolic generative system that allows you to express complex meanings, but it’s a self-contained system: The meanings are highly abstract and mostly relational, and not connected to the real world that we experience.”

Some other characteristics of natural languages, which are not shared by constructed languages, don’t seem to be necessary to generate a response in the language network.

“It doesn’t matter whether the language is created and shaped over time by a community of speakers, because these constructed languages are not,” Malik-Moraleda says. “It doesn’t matter how old they are, because conlangs that are just a decade old engage the same brain regions as natural languages that have been around for many hundreds of years.”

To further refine the features of language that activate the brain’s language network, Fedorenko’s lab is now planning to study how the brain responds to a conlang called Lojban, which was created by the Logical Language Group in the 1990s and was designed to prevent ambiguity of meanings and promote more efficient communication.

The research was funded by MIT’s McGovern Institute for Brain Research, Brain and Cognitive Sciences Department, the Simons Center for the Social Brain, the Frederick A. and Carole J. Middleton Career Development Professorship, and the U.S. National Institutes of Health.

Prototyping large structures with integrated electronics, like a chair that can monitor someone’s sitting posture, is typically a laborious and wasteful process.

One might need to fabricate multiple versions of the chair structure via 3D printing and laser cutting, generating a great deal of waste, before assembling the frame, grafting sensors and other fragile electronics onto it, and then wiring it up to create a working device.

If the prototype fails, the maker will likely have no choice but to discard it and go back to the drawing board.

MIT researchers have come up with a better way to iteratively design large and sturdy interactive structures. They developed a rapid development platform that utilizes reconfigurable building blocks with integrated electronics that can be assembled into complex, functional devices. Rather than building electronics into a structure, the electronics become the structure.

These lightweight three-dimensional lattice building blocks, known as voxels, have high strength and stiffness, along with integrated sensing, response, and processing abilities that enable users without mechanical or electrical engineering expertise to rapidly produce interactive electronic devices.

The voxels, which can be assembled, disassembled, and reconfigured almost infinitely into various forms, cost about 50 cents each.

The prototyping platform, called VIK (Voxel Invention Kit), includes a user-friendly design tool that enables end-to-end prototyping, allowing a user to simulate the structure’s response to mechanical loads and iterate on the design as needed.

“This is about democratizing access to functional interactive devices. With VIK, there is no 3D printing or laser cutting required. If you just have the voxel faces, you are able to produce these interactive structures anywhere you want,” says Jack Forman, an MIT graduate student in media arts and sciences and affiliate of the MIT Center for Bits and Atoms (CBA) and the MIT Media Lab, and co-lead author of a paper on VIK.

Forman is joined on the paper by co-lead author and fellow graduate student Miana Smith; graduate student Amira Abdel-Rahman; and senior author Neil Gershenfeld, an MIT professor and director of the CBA. The research will be presented at the Conference on Human Factors in Computing Systems.

Functional building blocks

VIK builds upon years of work in the CBA to develop discrete, cellular building blocks called voxels. One voxel, an aluminum cuboctahedra lattice (which has eight triangular faces and six square faces), is strong enough to support 228 kilograms, or about the weight of an upright piano.

Instead of being 3D printed, milled, or laser cut, voxels are assembled into largescale, strong, durable structures like airplane components or wind turbines that can respond to their environments.

The CBA team merged voxels other work in their lab centered on interconnected electrical components, yielding voxels with structural electronics. Assembling these functional voxels generates structures that can transmit data and power, as well as mechanical forces, without the need for wires.

They used these electromechanical building blocks to develop VIK.

“It was an interesting challenge to think about adapting a lot of our previous work, which has been about hitting hard engineering metrics, into a user-friendly system that makes sense and is fun and easy for people to work with,” Smith says.

For instance, they made the voxel design larger so the lattice structures are easier for human hands to assemble and disassemble. They also added aluminum cross-bracing to the units to improve their strength and stability.

In addition, VIK voxels have a reversible, snap-fit connection so a user can seamlessly assemble them without the need for additional tools, in contrast to some prior voxel designs that used rivets as fasteners.

“We designed the voxel faces to permit only the correct connections. That means that, if you are building with voxels, you are guaranteed to be building the correct wiring harness. Once you finish your device, you can just plug it in and it will work,” says Smith.

Wiring harnesses can add significant cost to functional systems and can often be a source of failure.

An accessible prototyping platform

To help users who have minimal engineering expertise create a wide array of interactive devices, the team developed a user-friendly interface to simulate 3D voxel structures.

The interface includes a Finite Element Analysis (FEA) simulation model that enables users to draw out a structure and simulate the forces and mechanical loads that will be applied to it. It adds colors to an animation of the user’s device to identify potential points of failure.

“We created what is essentially a ‘Minecraft’ for voxel applications. You don’t need a good sense of civil engineering or truss analysis to verify that the structure you are making is safe. Anyone can build something with VIK and have confidence in it,” Forman says.

Users can also easily integrate off-the-shelf modules, like speakers, sensors, or actuators, into their device. VIK emphasizes flexibility, enabling makers to use the types of microcontrollers they are comfortable with.

“The next evolution of electronics will be in three-dimensional space and the Voxel Invention Kit (VIK) is the stepping stone that will enable users, designers, and innovators a way to visualize and integrate electronics directly into structures,” says Victor Zaderej, manager of advanced electronics packaging technology at Molex, a manufacturer of electronic, electrical, and fiber optic connectivity systems. “Think of the VIK as the merging of a LEGO building kit and an electronics breadboard. When creative engineers and designers begin thinking about potential applications, the opportunities and unique products that will be enabled will be limitless.”

Using the design tool for feedback, a maker can rapidly change the configuration of voxels to adjust a prototype or disassemble the structure to build something new. If the user eventually wishes to discard the device, the aluminum voxels are fully recyclable.

This reconfigurability and recyclability, along with the high strength, high stiffness, light weight, and integrated electronics of the voxels, could make VIK especially well-suited for applications such as theatrical stage design, where stage managers want to support actors safely with customizable set pieces that might only exist for a few days.

And by enabling the rapid-prototyping of large, complex structures, VIK could also have future applications in areas like space fabrication or in the development of smart buildings and intelligent infrastructure for sustainable cities.

But for the researchers, perhaps the most important next step will be to get VIK out into the world to see what users come up with.

“These voxels are now so readily available that someone can use them in their day-to-day life. It will be exciting to see what they can do and create with VIK,” adds Forman.

We move thanks to coordination among many skeletal muscle fibers, all twitching and pulling in sync. While some muscles align in one direction, others form intricate patterns, helping parts of the body move in multiple ways.

In recent years, scientists and engineers have looked to muscles as potential actuators for “biohybrid” robots — machines powered by soft, artificially grown muscle fibers. Such bio-bots could squirm and wiggle through spaces where traditional machines cannot. For the most part, however, researchers have only been able to fabricate artificial muscle that pulls in one direction, limiting any robot’s range of motion.

Now MIT engineers have developed a method to grow artificial muscle tissue that twitches and flexes in multiple coordinated directions. As a demonstration, they grew an artificial, muscle-powered structure that pulls both concentrically and radially, much like how the iris in the human eye acts to dilate and constrict the pupil.

The researchers fabricated the artificial iris using a new “stamping” approach they developed. First, they 3D-printed a small, handheld stamp patterned with microscopic grooves, each as small as a single cell. Then they pressed the stamp into a soft hydrogel and seeded the resulting grooves with real muscle cells. The cells grew along these grooves within the hydrogel, forming fibers. When the researchers stimulated the fibers, the muscle contracted in multiple directions, following the fibers’ orientation.

“With the iris design, we believe we have demonstrated the first skeletal muscle-powered robot that generates force in more than one direction. That was uniquely enabled by this stamp approach,” says Ritu Raman, the Eugene Bell Career Development Professor of Tissue Engineering in MIT’s Department of Mechanical Engineering.

The team says the stamp can be printed using tabletop 3D printers and fitted with different patterns of microscopic grooves. The stamp can be used to grow complex patterns of muscle — and potentially other types of biological tissues, such as neurons and heart cells — that look and act like their natural counterparts.

“We want to make tissues that replicate the architectural complexity of real tissues,” Raman says. “To do that, you really need this kind of precision in your fabrication.”

She and her colleagues published their open-access results Friday in the journal Biomaterials Science. Her MIT co-authors include first author Tamara Rossy, Laura Schwendeman, Sonika Kohli, Maheera Bawa, and Pavankumar Umashankar, along with Roi Habba, Oren Tchaicheeyan, and Ayelet Lesman of Tel Aviv University in Israel.

Training space

Raman’s lab at MIT aims to engineer biological materials that mimic the sensing, activity, and responsiveness of real tissues in the body. Broadly, her group seeks to apply these bioengineered materials in areas from medicine to machines. For instance, she is looking to fabricate artificial tissue that can restore function to people with neuromuscular injury. She is also exploring artificial muscles for use in soft robotics, such as muscle-powered swimmers that move through the water with fish-like flexibility.

Raman has previously developed what could be seen as gym platforms and workout routines for lab-grown muscle cells. She and her colleagues designed a hydrogel “mat” that encourages muscle cells to grow and fuse into fibers without peeling away. She also derived a way to “exercise” the cells by genetically engineering them to twitch in response to pulses of light. And, her group has come up with ways to direct muscle cells to grow in long, parallel lines, similar to natural, striated muscles. However, it’s been a challenge, for her group and others, to design artificial muscle tissue that moves in multiple, predictable directions.

“One of the cool things about natural muscle tissues is, they don’t just point in one direction. Take for instance, the circular musculature in our iris and around our trachea. And even within our arms and legs, muscle cells don’t point straight, but at an angle,” Raman notes. “Natural muscle has multiple orientations in the tissue, but we haven’t been able to replicate that in our engineered muscles.”

Muscle blueprint

In thinking of ways to grow multidirectional muscle tissue, the team hit on a surprisingly simple idea: stamps. Inspired in part by the classic Jell-O mold, the team looked to design a stamp, with microscopic patterns that could be imprinted into a hydrogel, similar to the muscle-training mats that the group has previously developed. The patterns of the imprinted mat could then serve as a roadmap along which muscle cells might follow and grow.

“The idea is simple. But how do you make a stamp with features as small as a single cell? And how do you stamp something that’s super soft? This gel is much softer than Jell-O, and it’s something that’s really hard to cast, because it could tear really easily,” Raman says.

The team tried variations on the stamp design and eventually landed on an approach that worked surprisingly well. The researchers fabricated a small, handheld stamp using high-precision printing facilities in MIT.nano, which enabled them to print intricate patterns of grooves, each about as wide as a single muscle cell, onto the bottom of the stamp. Before pressing the stamp into a hydrogel mat, they coated the bottom with a protein that helped the stamp imprint evenly into the gel and peel away without sticking or tearing.

As a demonstration, the researchers printed a stamp with a pattern similar to the microscopic musculature in the human iris. The iris comprises a ring of muscle surrounding the pupil. This ring of muscle is made up of an inner circle of muscle fibers arranged concentrically, following a circular pattern, and an outer circle of fibers that stretch out radially, like the rays of the sun.  Together, this complex architecture acts to constrict or dilate the pupil.

Once Raman and her colleagues pressed the iris pattern into a hydrogel mat, they coated the mat with cells that they genetically engineered to respond to light. Within a day, the cells fell into the microscopic grooves and began to fuse into fibers, following the iris-like patterns and eventually growing into a whole muscle, with an architecture and size similar to a real iris.

When the team stimulated the artificial iris with pulses of light, the muscle contracted in multiple directions, similar to the iris in the human eye. Raman notes that the team’s artificial iris is fabricated with skeletal muscle cells, which are involved in voluntary motion, whereas the muscle tissue in the real human iris is made up of smooth muscle cells, which are a type of involuntary muscle tissue. They chose to pattern skeletal muscle cells in an iris-like pattern to demonstrate the ability to fabricate complex, multidirectional muscle tissue.

“In this work, we wanted to show we can use this stamp approach to make a ‘robot’ that can do things that previous muscle-powered robots can’t do,” Raman says. “We chose to work with skeletal muscle cells. But there’s nothing stopping you from doing this with any other cell type.”

She notes that while the team used precision-printing techniques, the stamp design can also be made using conventional tabletop 3D printers. Going forward, she and her colleagues plan to apply the stamping method to other cell types, as well as explore different muscle architectures and ways to activate artificial, multidirectional muscle to do useful work.

“Instead of using rigid actuators that are typical in underwater robots, if we can use soft biological robots, we can navigate and be much more energy-efficient, while also being completely biodegradable and sustainable,” Raman says. “That’s what we hope to build toward.”

This work was supported, in part, by the U.S. Office of Naval Research, the U.S. Army Research Office, the U.S. National Science Foundation, and the U.S. National Institutes of Health.

A decade after scientists in The Picower Institute for Learning and Memory at MIT first began testing whether sensory stimulation of the brain’s 40Hz “gamma” frequency rhythms could treat Alzheimer’s disease in mice, a growing evidence base supporting the idea that it can improve brain health — in humans as well as animals — has emerged from the work of labs all over the world. A new open-access review article in PLOS Biology describes the state of research so far and presents some of the fundamental and clinical questions at the forefront of the noninvasive gamma stimulation now.

“As we’ve made all our observations, many other people in the field have published results that are very consistent,” says Li-Huei Tsai, Picower professor of neuroscience at MIT, director of MIT’s Aging Brain Initiative, and senior author of the new review, with postdoc Jung Park. “People have used many different ways to induce gamma including sensory stimulation, transcranial alternating current stimulation, or transcranial magnetic stimulation, but the key is delivering stimulation at 40 hertz. They all see beneficial effects.”

A decade of discovery at MIT

Starting with a paper in Nature in 2016, a collaboration led by Tsai has produced a series of studies showing that 40Hz stimulation via light, sound, the two combined, or tactile vibration reduces hallmarks of Alzheimer’s pathology such as amyloid and tau proteins, prevents neuron death, decreases synapse loss, and sustains memory and cognition in various Alzheimer’s mouse models. The collaboration’s investigations of the underlying mechanisms that produce these benefits have so far identified specific cellular and molecular responses in many brain cell types including neurons, microglia, astrocytes, oligodendrocytes, and the brain’s blood vessels. Last year, for instance, the lab reported in Nature that 40Hz audio and visual stimulation induced interneurons in mice to increase release of the peptide VIP, prompting increased clearance of amyloid from brain tissue via the brain’s glymphatic “plumbing” system.

Meanwhile, at MIT and at the MIT spinoff company Cognito Therapeutics, phase II clinical studies have shown that people with Alzheimer’s exposed to 40Hz light and sound experienced a significant slowing of brain atrophy and improvements on some cognitive measures, compared to untreated controls. Cognito, which has also measured significant preservation of the brain’s “white matter” in volunteers, has been conducting a pivotal, nationwide phase III clinical trial of sensory gamma stimulation for more than a year.

“Neuroscientists often lament that it is a great time to have AD [Alzheimer’s disease] if you are a mouse,” Park and Tsai wrote in the review. “Our ultimate goal, therefore, is to translate GENUS discoveries into a safe, accessible, and noninvasive therapy for AD patients.” The MIT team often refers to 40Hz stimulation as “GENUS” for Gamma Entrainment Using Sensory Stimulation.

A growing field

As Tsai’s collaboration, which includes MIT colleagues Edward Boyden and Emery N. Brown, has published its results, many other labs have produced studies adding to the evidence that various methods of noninvasive gamma sensory stimulation can combat Alzheimer’s pathology. Among many examples cited in the new review, in 2024 a research team in China independently corroborated that 40Hz sensory stimulation increases glymphatic fluid flows in mice. In another example, a Harvard Medical School-based team in 2022 showed that 40Hz gamma stimulation using Transcranial Alternating Current Stimulation significantly reduced the burden of tau in three out of four human volunteers. And in another study involving more than 100 people, researchers in Scotland in 2023 used audio and visual gamma stimulation (at 37.5Hz) to improve memory recall.

Open questions

Amid the growing number of publications describing preclinical studies with mice and clinical trials with people, open questions remain, Tsai and Park acknowledge. The MIT team and others are still exploring the cellular and molecular mechanisms that underlie GENUS’s effects. Tsai says her lab is looking at other neuropeptide and neuromodulatory systems to better understand the cascade of events linking sensory stimulation to the observed cellular responses. Meanwhile, the nature of how some cells, such as microglia, respond to gamma stimulation and how that affects pathology remains unclear, Tsai adds.

Even with a national phase III clinical trial underway, it is still important to investigate these fundamental mechanisms, Tsai says, because new insights into how noninvasive gamma stimulation affects the brain could improve and expand its therapeutic potential.

“The more we understand the mechanisms, the more we will have good ideas about how to further optimize the treatment,” Tsai says. “And the more we understand its action and the circuits it affects, the more we will know beyond Alzheimer’s disease what other neurological disorders will benefit from this.”

Indeed, the review points to studies at MIT and other institutions providing at least some evidence that GENUS might be able to help with Parkinson’s disease, stroke, anxiety, epilepsy, and the cognitive side effects of chemotherapy and conditions that reduce myelin, such as multiple sclerosis. Tsai’s lab has been studying whether it can help with Down syndrome as well.

The open questions may help define the next decade of GENUS research.

It is a deep question, from deep in our history: When did human language as we know it emerge? A new survey of genomic evidence suggests our unique language capacity was present at least 135,000 years ago. Subsequently, language might have entered social use 100,000 years ago.

Our species, Homo sapiens, is about 230,000 years old. Estimates of when language originated vary widely, based on different forms of evidence, from fossils to cultural artifacts. The authors of the new analysis took a different approach. They reasoned that since all human languages likely have a common origin — as the researchers strongly think — the key question is how far back in time regional groups began spreading around the world.

“The logic is very simple,” says Shigeru Miyagawa, an MIT professor and co-author of a new paper summarizing the results. “Every population branching across the globe has human language, and all languages are related.” Based on what the genomics data indicate about the geographic divergence of early human populations, he adds, “I think we can say with a fair amount of certainty that the first split occurred about 135,000 years ago, so human language capacity must have been present by then, or before.”

The paper, “Linguistic capacity was present in the Homo sapiens population 135 thousand years ago,” appears in Frontiers in Psychology. The co-authors are Miyagawa, who is a professor emeritus of linguistics and the Kochi-Manjiro Professor of Japanese Language and Culture at MIT; Rob DeSalle, a principal investigator at the American Museum of Natural History’s Institute for Comparative Genomics; Vitor Augusto Nóbrega, a faculty member in linguistics at the University of São Paolo; Remo Nitschke, of the University of Zurich, who worked on the project while at the University of Arizona linguistics department; Mercedes Okumura of the Department of Genetics and Evolutionary Biology at the University of São Paulo; and Ian Tattersall, curator emeritus of human origins at the American Museum of Natural History.

The new paper examines 15 genetic studies of different varieties, published over the past 18 years: Three used data about the inherited Y chromosome, three examined mitochondrial DNA, and nine were whole-genome studies.

All told, the data from these studies suggest an initial regional branching of humans about 135,000 years ago. That is, after the emergence of Homo sapiens, groups of people subsequently moved apart geographically, and some resulting genetic variations have developed, over time, among the different regional subpopulations. The amount of genetic variation shown in the studies allows researchers to estimate the point in time at which Homo sapiens was still one regionally undivided group.

Miyagawa says the studies collectively provide increasingly converging evidence about when these geographic splits started taking place. The first survey of this type was performed by other scholars in 2017, but they had fewer existing genetic studies to draw upon. Now, there are much more published data available, which when considered together point to 135,000 years ago as the likely time of the first split.

The new meta-analysis was possible because “quantity-wise we have more studies, and quality-wise, it’s a narrower window [of time],” says Miyagawa, who also holds an appointment at the University of São Paolo.

Like many linguists, Miyagawa believes all human languages are demonstrably related to each other, something he has examined in his own work. For instance, in his 2010 book, “Why Agree? Why Move?” he analyzed previously unexplored similarities between English, Japanese, and some of the Bantu languages. There are more than 7,000 identified human languages around the globe.

Some scholars have proposed that language capacity dates back a couple of million years, based on the physiological characteristics of other primates. But to Miyagawa, the question is not when primates could utter certain sounds; it is when humans had the cognitive ability to develop language as we know it, combining vocabulary and grammar into a system generating an infinite amount of rules-based expression.

“Human language is qualitatively different because there are two things, words and syntax, working together to create this very complex system,” Miyagawa says. “No other animal has a parallel structure in their communication system. And that gives us the ability to generate very sophisticated thoughts and to communicate them to others.”

This conception of human language origins also holds that humans had the cognitive capacity for language for some period of time before we constructed our first languages.

“Language is both a cognitive system and a communication system,” Miyagawa says. “My guess is prior to 135,000 years ago, it did start out as a private cognitive system, but relatively quickly that turned into a communications system.”

So, how can we know when distinctively human language was first used? The archaeological record is invaluable in this regard. Roughly 100,000 years ago, the evidence shows, there was a widespread appearance of symbolic activity, from meaningful markings on objects to the use of fire to produce ochre, a decorative red color.

Like our complex, highly generative language, these symbolic activities are engaged in by people, and no other creatures. As the paper notes, “behaviors compatible with language and the consistent exercise of symbolic thinking are detectable only in the archaeological record of H. sapiens.”

Among the co-authors, Tattersall has most prominently propounded the view that language served as a kind of ignition for symbolic thinking and other organized activities.

“Language was the trigger for modern human behavior,” Miyagawa says. “Somehow it stimulated human thinking and helped create these kinds of behaviors. If we are right, people were learning from each other [due to language] and encouraging innovations of the types we saw 100,000 years ago.”

To be sure, as the authors acknowledge in the paper, other scholars believe there was a more incremental and broad-based development of new activities around 100,000 years ago, involving materials, tools, and social coordination, with language playing a role in this, but not necessarily being the central force.

For his part, Miyagawa recognizes that there is considerable room for further progress in this area of research, but thinks efforts like the current paper are at least steps toward filling out a more detailed picture of language’s emergence.

“Our approach is very empirically based, grounded in the latest genetic understanding of early homo sapiens,” Miyagawa says. “I think we are on a good research arc, and I hope this will encourage people to look more at human language and evolution.”

This research was, in part, supported by the São Paolo Excellence Chair awarded to Miyagawa by the São Paolo Research Foundation.

Many companies invest heavily in hiring talent to create the high-performance library code that underpins modern artificial intelligence systems. NVIDIA, for instance, developed some of the most advanced high-performance computing (HPC) libraries, creating a competitive moat that has proven difficult for others to breach.

But what if a couple of students, within a few months, could compete with state-of-the-art HPC libraries with a few hundred lines of code, instead of tens or hundreds of thousands?

That’s what researchers at MIT’s Computer Science and Artificial Intelligence Laboratory (CSAIL) have shown with a new programming language called Exo 2.

Exo 2 belongs to a new category of programming languages that MIT Professor Jonathan Ragan-Kelley calls “user-schedulable languages” (USLs). Instead of hoping that an opaque compiler will auto-generate the fastest possible code, USLs put programmers in the driver's seat, allowing them to write “schedules” that explicitly control how the compiler generates code. This enables performance engineers to transform simple programs that specify what they want to compute into complex programs that do the same thing as the original specification, but much, much faster.

One of the limitations of existing USLs (like the original Exo) is their relatively fixed set of scheduling operations, which makes it difficult to reuse scheduling code across different “kernels” (the individual components in a high-performance library).

In contrast, Exo 2 enables users to define new scheduling operations externally to the compiler, facilitating the creation of reusable scheduling libraries. Lead author Yuka Ikarashi, an MIT PhD student in electrical engineering and computer science and CSAIL affiliate, says that Exo 2 can reduce total schedule code by a factor of 100 and deliver performance competitive with state-of-the-art implementations on multiple different platforms, including Basic Linear Algebra Subprograms (BLAS) that power many machine learning applications. This makes it an attractive option for engineers in HPC focused on optimizing kernels across different operations, data types, and target architectures.

“It’s a bottom-up approach to automation, rather than doing an ML/AI search over high-performance code,” says Ikarashi. “What that means is that performance engineers and hardware implementers can write their own scheduling library, which is a set of optimization techniques to apply on their hardware to reach the peak performance.”

One major advantage of Exo 2 is that it reduces the amount of coding effort needed at any one time by reusing the scheduling code across applications and hardware targets. The researchers implemented a scheduling library with roughly 2,000 lines of code in Exo 2, encapsulating reusable optimizations that are linear-algebra specific and target-specific (AVX512, AVX2, Neon, and Gemmini hardware accelerators). This library consolidates scheduling efforts across more than 80 high-performance kernels with up to a dozen lines of code each, delivering performance comparable to, or better than, MKL, OpenBLAS, BLIS, and Halide.

Exo 2 includes a novel mechanism called “Cursors” that provides what they call a “stable reference” for pointing at the object code throughout the scheduling process. Ikarashi says that a stable reference is essential for users to encapsulate schedules within a library function, as it renders the scheduling code independent of object-code transformations.

“We believe that USLs should be designed to be user-extensible, rather than having a fixed set of operations,” says Ikarashi. “In this way, a language can grow to support large projects through the implementation of libraries that accommodate diverse optimization requirements and application domains.”

Exo 2’s design allows performance engineers to focus on high-level optimization strategies while ensuring that the underlying object code remains functionally equivalent through the use of safe primitives. In the future, the team hopes to expand Exo 2’s support for different types of hardware accelerators, like GPUs. Several ongoing projects aim to improve the compiler analysis itself, in terms of correctness, compilation time, and expressivity.

Ikarashi and Ragan-Kelley co-authored the paper with graduate students Kevin Qian and Samir Droubi, Alex Reinking of Adobe, and former CSAIL postdoc Gilbert Bernstein, now a professor at the University of Washington. This research was funded, in part, by the U.S. Defense Advanced Research Projects Agency (DARPA) and the U.S. National Science Foundation, while the first author was also supported by Masason, Funai, and Quad Fellowships.

Converting one type of cell to another — for example, a skin cell to a neuron — can be done through a process that requires the skin cell to be induced into a “pluripotent” stem cell, then differentiated into a neuron. Researchers at MIT have now devised a simplified process that bypasses the stem cell stage, converting a skin cell directly into a neuron.

Working with mouse cells, the researchers developed a conversion method that is highly efficient and can produce more than 10 neurons from a single skin cell. If replicated in human cells, this approach could enable the generation of large quantities of motor neurons, which could potentially be used to treat patients with spinal cord injuries or diseases that impair mobility.

“We were able to get to yields where we could ask questions about whether these cells can be viable candidates for the cell replacement therapies, which we hope they could be. That’s where these types of reprogramming technologies can take us,” says Katie Galloway, the W. M. Keck Career Development Professor in Biomedical Engineering and Chemical Engineering.

As a first step toward developing these cells as a therapy, the researchers showed that they could generate motor neurons and engraft them into the brains of mice, where they integrated with host tissue.

Galloway is the senior author of two papers describing the new method, which appear today in Cell Systems. MIT graduate student Nathan Wang is the lead author of both papers.

From skin to neurons

Nearly 20 years ago, scientists in Japan showed that by delivering four transcription factors to skin cells, they could coax them to become induced pluripotent stem cells (iPSCs). Similar to embryonic stem cells, iPSCs can be differentiated into many other cell types. This technique works well, but it takes several weeks, and many of the cells don’t end up fully transitioning to mature cell types.

“Oftentimes, one of the challenges in reprogramming is that cells can get stuck in intermediate states,” Galloway says. “So, we’re using direct conversion, where instead of going through an iPSC intermediate, we’re going directly from a somatic cell to a motor neuron.”

Galloway’s research group and others have demonstrated this type of direct conversion before, but with very low yields — fewer than 1 percent. In Galloway’s previous work, she used a combination of six transcription factors plus two other proteins that stimulate cell proliferation. Each of those eight genes was delivered using a separate viral vector, making it difficult to ensure that each was expressed at the correct level in each cell.

In the first of the new Cell Systems papers, Galloway and her students reported a way to streamline the process so that skin cells can be converted to motor neurons using just three transcription factors, plus the two genes that drive cells into a highly proliferative state.

Using mouse cells, the researchers started with the original six transcription factors and experimented with dropping them out, one at a time, until they reached a combination of three — NGN2, ISL1, and LHX3 — that could successfully complete the conversion to neurons.

Once the number of genes was down to three, the researchers could use a single modified virus to deliver all three of them, allowing them to ensure that each cell expresses each gene at the correct levels.

Using a separate virus, the researchers also delivered genes encoding p53DD and a mutated version of HRAS. These genes drive the skin cells to divide many times before they start converting to neurons, allowing for a much higher yield of neurons, about 1,100 percent.

“If you were to express the transcription factors at really high levels in nonproliferative cells, the reprogramming rates would be really low, but hyperproliferative cells are more receptive. It’s like they’ve been potentiated for conversion, and then they become much more receptive to the levels of the transcription factors,” Galloway says.

The researchers also developed a slightly different combination of transcription factors that allowed them to perform the same direct conversion using human cells, but with a lower efficiency rate — between 10 and 30 percent, the researchers estimate. This process takes about five weeks, which is slightly faster than converting the cells to iPSCs first and then turning them into neurons.

Implanting cells

Once the researchers identified the optimal combination of genes to deliver, they began working on the best ways to deliver them, which was the focus of the second Cell Systems paper.

They tried out three different delivery viruses and found that a retrovirus achieved the most efficient rate of conversion. Reducing the density of cells grown in the dish also helped to improve the overall yield of motor neurons. This optimized process, which takes about two weeks in mouse cells, achieved a yield of more than 1,000 percent.

Working with colleagues at Boston University, the researchers then tested whether these motor neurons could be successfully engrafted into mice. They delivered the cells to a part of the brain known as the striatum, which is involved in motor control and other functions.

After two weeks, the researchers found that many of the neurons had survived and seemed to be forming connections with other brain cells. When grown in a dish, these cells showed measurable electrical activity and calcium signaling, suggesting the ability to communicate with other neurons. The researchers now hope to explore the possibility of implanting these neurons into the spinal cord.

The MIT team also hopes to increase the efficiency of this process for human cell conversion, which could allow for the generation of large quantities of neurons that could be used to treat spinal cord injuries or diseases that affect motor control, such as ALS. Clinical trials using neurons derived from iPSCs to treat ALS are now underway, but expanding the number of cells available for such treatments could make it easier to test and develop them for more widespread use in humans, Galloway says.

The research was funded by the National Institute of General Medical Sciences and the National Science Foundation Graduate Research Fellowship Program.

For those looking to climb the corporate ladder in the U.S., here’s an idea you might not have considered: debate training.

According to a new research paper, people who learn the basics of debate are more likely to advance to leadership roles in U.S. organizations, compared to those who do not receive this training. One key reason is that being equipped with debate skills makes people more assertive in the workplace.

“Debate training can promote leadership emergence and advancement by fostering individuals’ assertiveness, which is a key, valued leadership characteristic in U.S. organizations,” says MIT Associate Professor Jackson Lu, one of the scholars who conducted the study.

The research is based on two experiments and provides empirical insights into leadership development, a subject more often discussed anecdotally than studied systematically.

“Leadership development is a multi-billion-dollar industry, where people spend a lot of money trying to help individuals emerge as leaders,” Lu says. “But the public doesn’t actually know what would be effective, because there hasn’t been a lot of causal evidence. That’s exactly what we provide.”

The paper, “Breaking Ceilings: Debate Training Promotes Leadership Emergence by Increasing Assertiveness,” was published Monday in the Journal of Applied Psychology. The authors are Lu, an associate professor at the MIT Sloan School of Management; Michelle X. Zhao, an undergraduate student at the Olin Business School of Washington University in St. Louis; Hui Liao, a professor and assistant dean at the University of Maryland’s Robert H. Smith School of Business; and Lu Doris Zhang, a doctoral student at MIT Sloan.

Assertiveness in the attention economy

The researchers conducted two experiments. In the first, 471 employees in a Fortune 100 firm were randomly assigned to receive either nine weeks of debate training or no training. Examined 18 months later, those receiving debate training were more likely to have advanced to leadership roles, by about 12 percentage points. This effect was statistically explained by increased assertiveness among those with debate training.

The second experiment, conducted with 975 university participants, further tested the causal effects of debate training in a controlled setting. Participants were randomly assigned to receive debate training, an alternative non-debate training, or no training. Consistent with the first experiment, participants receiving the debate training were more likely to emerge as leaders in subsequent group activities, an effect statistically explained by their increased assertiveness.

“The inclusion of a non-debate training condition allowed us to causally claim that debate training, rather than just any training, improved assertiveness and increased leadership emergence,” Zhang says. 

To some people, increasing assertiveness might not seem like an ideal recipe for success in an organizational setting, as it might seem likely to increase tensions or decrease cooperation. But as the authors note, the American Psychological Association conceptualizes assertiveness as “an adaptive style of communication in which individuals express their feelings and needs directly, while maintaining respect for others.”

Lu adds: “Assertiveness is conceptually different from aggressiveness. To speak up in meetings or classrooms, people don’t need to be aggressive jerks. You can ask questions politely, yet still effectively express opinons. Of course, that’s different from not saying anything at all.”

Moreover, in the contemporary world where we all must compete for attention, refined communication skills may be more important than ever.

“Whether it is cutting filler or mastering pacing, knowing how to assert our opinions helps us sound more leader-like,” Zhang says.

How firms identify leaders

The research also finds that debate training benefits people across demographics: Its impact was not significantly different for men or women, for those born in the U.S. or outside it, or for different ethnic groups.

However, the findings raise still other questions about how firms identify leaders. As the results show, individuals might have incentive to seek debate training and other general workplace skills. But how much responsibility do firms have to understand and recognize the many kinds of skills, beyond assertiveness, that employees may have?

“We emphasize that the onus of breaking leadership barriers should not fall on individuals themelves,” Lu says. “Organizations should also recognize and appreciate different communication and leadership styles in the workplace.”

Lu also notes that ongoing work is needed to understand if those firms are properly valuing the attributes of their own leaders.

“There is an important distinction between leadership emergence and leadership effectiveness,” Lu says. “Our paper looks at leadership emergence. It’s possible that people who are better listeners, who are more cooperative, and humbler, should also be selected for leadership positions because they are more effective leaders.”

This research was partly funded by the Society for Personality and Social Psychology.

Look around, and you’ll see it everywhere: the way trees form branches, the way cities divide into neighborhoods, the way the brain organizes into regions. Nature loves modularity — a limited number of self-contained units that combine in different ways to perform many functions. But how does this organization arise? Does it follow a detailed genetic blueprint, or can these structures emerge on their own?

A new study from MIT Professor Ila Fiete suggests a surprising answer.

In findings published Feb. 18 in Nature, Fiete, an associate investigator in the McGovern Institute for Brain Research and director of the K. Lisa Yang Integrative Computational Neuroscience (ICoN) Center at MIT, reports that a mathematical model called peak selection can explain how modules emerge without strict genetic instructions. Her team’s findings, which apply to brain systems and ecosystems, help explain how modularity occurs across nature, no matter the scale.

Joining two big ideas

“Scientists have debated how modular structures form. One hypothesis suggests that various genes are turned on at different locations to begin or end a structure. This explains how insect embryos develop body segments, with genes turning on or off at specific concentrations of a smooth chemical gradient in the insect egg,” says Fiete, who is the senior author of the paper. Mikail Khona PhD '25, a former graduate student and K. Lisa Yang ICoN Center graduate fellow, and postdoc Sarthak Chandra also led the study.

Another idea, inspired by mathematician Alan Turing, suggests that a structure could emerge from competition — small-scale interactions can create repeating patterns, like the spots on a cheetah or the ripples in sand dunes.

Both ideas work well in some cases, but fail in others. The new research suggests that nature need not pick one approach over the other. The authors propose a simple mathematical principle called peak selection, showing that when a smooth gradient is paired with local interactions that are competitive, modular structures emerge naturally. “In this way, biological systems can organize themselves into sharp modules without detailed top-down instruction,” says Chandra.

Modular systems in the brain

The researchers tested their idea on grid cells, which play a critical role in spatial navigation as well as the storage of episodic memories. Grid cells fire in a repeating triangular pattern as animals move through space, but they don’t all work at the same scale — they are organized into distinct modules, each responsible for mapping space at slightly different resolutions.

No one knows how these modules form, but Fiete’s model shows that gradual variations in cellular properties along one dimension in the brain, combined with local neural interactions, could explain the entire structure. The grid cells naturally sort themselves into distinct groups with clear boundaries, without external maps or genetic programs telling them where to go. “Our work explains how grid cell modules could emerge. The explanation tips the balance toward the possibility of self-organization. It predicts that there might be no gene or intrinsic cell property that jumps when the grid cell scale jumps to another module,” notes Khona.

Modular systems in nature

The same principle applies beyond neuroscience. Imagine a landscape where temperatures and rainfall vary gradually over a space. You might expect species to be spread, and also to vary, smoothly over this region. But in reality, ecosystems often form species clusters with sharp boundaries — distinct ecological “neighborhoods” that don’t overlap.

Fiete’s study suggests why: local competition, cooperation, and predation between species interact with the global environmental gradients to create natural separations, even when the underlying conditions change gradually. This phenomenon can be explained using peak selection — and suggests that the same principle that shapes brain circuits could also be at play in forests and oceans.

A self-organizing world

One of the researchers’ most striking findings is that modularity in these systems is remarkably robust. Change the size of the system, and the number of modules stays the same — they just scale up or down. That means a mouse brain and a human brain could use the same fundamental rules to form their navigation circuits, just at different sizes.

The model also makes testable predictions. If it’s correct, grid cell modules should follow simple spacing ratios. In ecosystems, species distributions should form distinct clusters even without sharp environmental shifts.

Fiete notes that their work adds another conceptual framework to biology. “Peak selection can inform future experiments, not only in grid cell research but across developmental biology.”

MIT aerospace engineers have found that greenhouse gas emissions are changing the environment of near-Earth space in ways that, over time, will reduce the number of satellites that can sustainably operate there.

In a study appearing today in Nature Sustainability, the researchers report that carbon dioxide and other greenhouse gases can cause the upper atmosphere to shrink. An atmospheric layer of special interest is the thermosphere, where the International Space Station and most satellites orbit today. When the thermosphere contracts, the decreasing density reduces atmospheric drag — a force that pulls old satellites and other debris down to altitudes where they will encounter air molecules and burn up.

Less drag therefore means extended lifetimes for space junk, which will litter sought-after regions for decades and increase the potential for collisions in orbit.

The team carried out simulations of how carbon emissions affect the upper atmosphere and orbital dynamics, in order to estimate the “satellite carrying capacity” of low Earth orbit. These simulations predict that by the year 2100, the carrying capacity of the most popular regions could be reduced by 50-66 percent due to the effects of greenhouse gases.

“Our behavior with greenhouse gases here on Earth over the past 100 years is having an effect on how we operate satellites over the next 100 years,” says study author Richard Linares, associate professor in MIT’s Department of Aeronautics and Astronautics (AeroAstro).

“The upper atmosphere is in a fragile state as climate change disrupts the status quo,” adds lead author William Parker, a graduate student in AeroAstro. “At the same time, there’s been a massive increase in the number of satellites launched, especially for delivering broadband internet from space. If we don’t manage this activity carefully and work to reduce our emissions, space could become too crowded, leading to more collisions and debris.”

The study includes co-author Matthew Brown of the University of Birmingham.

Sky fall

The thermosphere naturally contracts and expands every 11 years in response to the sun’s regular activity cycle. When the sun’s activity is low, the Earth receives less radiation, and its outermost atmosphere temporarily cools and contracts before expanding again during solar maximum.

In the 1990s, scientists wondered what response the thermosphere might have to greenhouse gases. Their preliminary modeling showed that, while the gases trap heat in the lower atmosphere, where we experience global warming and weather, the same gases radiate heat at much higher altitudes, effectively cooling the thermosphere. With this cooling, the researchers predicted that the thermosphere should shrink, reducing atmospheric density at high altitudes.

In the last decade, scientists have been able to measure changes in drag on satellites, which has provided some evidence that the thermosphere is contracting in response to something more than the sun’s natural, 11-year cycle.

“The sky is quite literally falling — just at a rate that’s on the scale of decades,” Parker says. “And we can see this by how the drag on our satellites is changing.”

The MIT team wondered how that response will affect the number of satellites that can safely operate in Earth’s orbit. Today, there are over 10,000 satellites drifting through low Earth orbit, which describes the region of space up to 1,200 miles (2,000 kilometers), from Earth’s surface. These satellites deliver essential services, including internet, communications, navigation, weather forecasting, and banking. The satellite population has ballooned in recent years, requiring operators to perform regular collision-avoidance maneuvers to keep safe. Any collisions that do occur can generate debris that remains in orbit for decades or centuries, increasing the chance for follow-on collisions with satellites, both old and new.

“More satellites have been launched in the last five years than in the preceding 60 years combined,” Parker says. “One of key things we’re trying to understand is whether the path we’re on today is sustainable.”

Crowded shells

In their new study, the researchers simulated different greenhouse gas emissions scenarios over the next century to investigate impacts on atmospheric density and drag. For each “shell,” or altitude range of interest, they then modeled the orbital dynamics and the risk of satellite collisions based on the number of objects within the shell. They used this approach to identify each shell’s “carrying capacity” — a term that is typically used in studies of ecology to describe the number of individuals that an ecosystem can support.

“We’re taking that carrying capacity idea and translating it to this space sustainability problem, to understand how many satellites low Earth orbit can sustain,” Parker explains.

The team compared several scenarios: one in which greenhouse gas concentrations remain at their level from the year 2000 and others where emissions change according to the Intergovernmental Panel on Climate Change (IPCC) Shared Socioeconomic Pathways (SSPs). They found that scenarios with continuing increases in emissions would lead to a significantly reduced carrying capacity throughout low Earth orbit.

In particular, the team estimates that by the end of this century, the number of satellites safely accommodated within the altitudes of 200 and 1,000 kilometers could be reduced by 50 to 66 percent compared with a scenario in which emissions remain at year-2000 levels. If satellite capacity is exceeded, even in a local region, the researchers predict that the region will experience a “runaway instability,” or a cascade of collisions that would create so much debris that satellites could no longer safely operate there.

Their predictions forecast out to the year 2100, but the team says that certain shells in the atmosphere today are already crowding up with satellites, particularly from recent “megaconstellations” such as SpaceX’s Starlink, which comprises fleets of thousands of small internet satellites.

“The megaconstellation is a new trend, and we’re showing that because of climate change, we’re going to have a reduced capacity in orbit,” Linares says. “And in local regions, we’re close to approaching this capacity value today.”

“We rely on the atmosphere to clean up our debris. If the atmosphere is changing, then the debris environment will change too,” Parker adds. “We show the long-term outlook on orbital debris is critically dependent on curbing our greenhouse gas emissions.”

This research is supported, in part, by the U.S. National Science Foundation, the U.S. Air Force, and the U.K. Natural Environment Research Council.

Tuberculosis lives and thrives in the lungs. When the bacteria that cause the disease are coughed into the air, they are thrust into a comparatively hostile environment, with drastic changes to their surrounding pH and chemistry. How these bacteria survive their airborne journey is key to their persistence, but very little is known about how they protect themselves as they waft from one host to the next.

Now MIT researchers and their collaborators have discovered a family of genes that becomes essential for survival specifically when the pathogen is exposed to the air, likely protecting the bacterium during its flight.

Many of these genes were previously considered to be nonessential, as they didn’t seem to have any effect on the bacteria’s role in causing disease when injected into a host. The new work suggests that these genes are indeed essential, though for transmission rather than proliferation.

“There is a blind spot that we have toward airborne transmission, in terms of how a pathogen can survive these sudden changes as it circulates in the air,” says Lydia Bourouiba, who is the head of the Fluid Dynamics of Disease Transmission Laboratory, an associate professor of civil and environmental engineering and mechanical engineering, and a core faculty member in the Instiute for Medical Engineering and Science at MIT. “Now we have a sense, through these genes, of what tools tuberculosis uses to protect itself.”

The team’s results, appearing this week in the Proceedings of the National Academy of Sciences, could provide new targets for tuberculosis therapies that simultaneously treat infection and prevent transmission.

“If a drug were to target the product of these same genes, it could effectively treat an individual, and even before that person is cured, it could keep the infection from spreading to others,” says Carl Nathan, chair of the Department of Microbiology and Immunology and R.A. Rees Pritchett Professor of Microbiology at Weill Cornell Medicine.

Nathan and Bourouiba are co-senior authors of the study, which includes MIT co-authors and mentees of Bourouiba in the Fluids and Health Network: co-lead author postdoc Xiaoyi Hu, postdoc Eric Shen, and student mentees Robin Jahn and Luc Geurts. The study also includes collaborators from Weill Cornell Medicine, the University of California at San Diego, Rockefeller University, Hackensack Meridian Health, and the University of Washington.

Pathogen’s perspective

Tuberculosis is a respiratory disease caused by Mycobacterium tuberculosis, a bacterium that most commonly affects the lungs and is transmitted through droplets that an infected individual expels into the air, often through coughing or sneezing. Tuberculosis is the single leading cause of death from infection, except during the major global pandemics caused by viruses.

“In the last 100 years, we have had the 1918 influenza, the 1981 HIV AIDS epidemic, and the 2019 SARS Cov2 pandemic,” Nathan notes. “Each of those viruses has killed an enormous number of people. And as they have settled down, we are left with a ‘permanent pandemic’ of tuberculosis.”

Much of the research on tuberculosis centers on its pathophysiology — the mechanisms by which the bacteria take over and infect a host — as well as ways to diagnose and treat the disease. For their new study, Nathan and Bourouiba focused on transmission of tuberculosis, from the perspective of the bacterium itself, to investigate what defenses it might rely on to help it survive its airborne transmission.

“This is one of the first attempts to look at tuberculosis from the airborne perspective, in terms of what is happening to the organism, at the level of being protected from these sudden changes and very harsh biophysical conditions,” Bourouiba says.

Critical defense

At MIT, Bourouiba studies the physics of fluids and the ways in which droplet dynamics can spread particles and pathogens. She teamed up with Nathan, who studies tuberculosis, and the genes that the bacteria rely on throughout their life cycle.

To get a handle on how tuberculosis can survive in the air, the team aimed to mimic the conditions that the bacterium experiences during transmission. The researchers first looked to develop a fluid that is similar in viscosity and droplet sizes to what a patient would cough or sneeze out into the air. Bourouiba notes that much of the experimental work that has been done on tuberculosis in the past has been based on a liquid solution that scientists use to grow the bacteria. But the team found that this liquid has a chemical composition that is very different from the fluid that tuberculosis patients actually cough and sneeze into the air.

Additionally, Bourouiba notes that fluid commonly sampled from tuberculosis patients is based on sputum that a patient spits out, for instance for a diagnostic test. “The fluid is thick and gooey and it’s what most of the tuberculosis world considers to represent what is happening in the body,” she says. “But it’s extraordinarily inefficient in spreading to others because it’s too sticky to break into inhalable droplets.”

Through Bourouiba’s work with fluid and droplet physics, the team determined the more realistic viscosity and likely size distribution of tuberculosis-carrying microdroplets that would be transmitted through the air. The team also characterized the droplet compositions, based on analyses of patient samples of infected lung tissues. They then created a more realistic fluid, with a composition, viscosity, surface tension and droplet size that is similar to what would be released into the air from exhalations.

Then, the researchers deposited different fluid mixtures onto plates in tiny individual droplets and measured in detail how they evaporate and what internal structure they leave behind. They observed that the new fluid tended to shield the bacteria at the center of the droplet as the droplet evaporated, compared to conventional fluids where bacteria tended to be more exposed to the air. The more realistic fluid was also capable of retaining more water.

Additionally, the team infused each droplet with bacteria containing genes with various knockdowns, to see whether the absence of certain genes would affect the bacteria’s survival as the droplets evaporated.

In this way, the team assessed the activity of over 4,000 tuberculosis genes and discovered a family of several hundred genes that seemed to become important specifically as the bacteria adapted to airborne conditions. Many of these genes are involved in repairing damage to oxidized proteins, such as proteins that have been exposed to air. Other activated genes have to do with destroying damaged proteins that are beyond repair.

“What we turned up was a candidate list that’s very long,” Nathan says. “There are hundreds of genes, some more prominently implicated than others, that may be critically involved in helping tuberculosis survive its transmission phase.”

The team acknowledges the experiments are not a complete analog of the bacteria’s biophysical transmission. In reality, tuberculosis is carried in droplets that fly through the air, evaporating as they go. In order to carry out their genetic analyses, the team had to work with droplets sitting on a plate. Under these constraints, they mimicked the droplet transmission as best they could, by setting the plates in an extremely dry chamber to accelerate the droplets’ evaporation, analogous to what they would experience in flight.

Going forward, the researchers have started experimenting with platforms that allow them to study the droplets in flight, in a range of conditions. They plan to focus on the new family of genes in even more realistic experiments, to confirm whether the genes do indeed shield Mycobacterium tuberculosis as it is transmitted through the air, potentially opening the way to weakening its airborne defenses.

“The idea of waiting to find someone with tuberculosis, then treating and curing them, is a totally inefficient way to stop the pandemic,” Nathan says. “Most people who exhale tuberculosis do not yet have a diagnosis. So we have to interrupt its transmission. And how do you do that, if you don’t know anything about the process itself? We have some ideas now.”

This work was supported, in part, by the National Institutes of Health, the Abby and Howard P. Milstein Program in Chemical Biology and Translational Medicine, and the Potts Memorial Foundation, the National Science Foundation Center for Analysis and Prediction of Pandemic Expansion (APPEX), Inditex, NASA Translational Research Institute for Space Health , and Analog Devices, Inc.

Imagine that a robot is helping you clean the dishes. You ask it to grab a soapy bowl out of the sink, but its gripper slightly misses the mark.

Using a new framework developed by MIT and NVIDIA researchers, you could correct that robot’s behavior with simple interactions. The method would allow you to point to the bowl or trace a trajectory to it on a screen, or simply give the robot’s arm a nudge in the right direction.

Unlike other methods for correcting robot behavior, this technique does not require users to collect new data and retrain the machine-learning model that powers the robot’s brain. It enables a robot to use intuitive, real-time human feedback to choose a feasible action sequence that gets as close as possible to satisfying the user’s intent.

When the researchers tested their framework, its success rate was 21 percent higher than an alternative method that did not leverage human interventions.

In the long run, this framework could enable a user to more easily guide a factory-trained robot to perform a wide variety of household tasks even though the robot has never seen their home or the objects in it.

“We can’t expect laypeople to perform data collection and fine-tune a neural network model. The consumer will expect the robot to work right out of the box, and if it doesn’t, they would want an intuitive mechanism to customize it. That is the challenge we tackled in this work,” says Felix Yanwei Wang, an electrical engineering and computer science (EECS) graduate student and lead author of a paper on this method.

His co-authors include Lirui Wang PhD ’24 and Yilun Du PhD ’24; senior author Julie Shah, an MIT professor of aeronautics and astronautics and the director of the Interactive Robotics Group in the Computer Science and Artificial Intelligence Laboratory (CSAIL); as well as Balakumar Sundaralingam, Xuning Yang, Yu-Wei Chao, Claudia Perez-D’Arpino PhD ’19, and Dieter Fox of NVIDIA. The research will be presented at the International Conference on Robots and Automation.

Mitigating misalignment

Recently, researchers have begun using pre-trained generative AI models to learn a “policy,” or a set of rules, that a robot follows to complete an action. Generative models can solve multiple complex tasks.

During training, the model only sees feasible robot motions, so it learns to generate valid trajectories for the robot to follow.

While these trajectories are valid, that doesn’t mean they always align with a user’s intent in the real world. The robot might have been trained to grab boxes off a shelf without knocking them over, but it could fail to reach the box on top of someone’s bookshelf if the shelf is oriented differently than those it saw in training.

To overcome these failures, engineers typically collect data demonstrating the new task and re-train the generative model, a costly and time-consuming process that requires machine-learning expertise.

Instead, the MIT researchers wanted to allow users to steer the robot’s behavior during deployment when it makes a mistake.

But if a human interacts with the robot to correct its behavior, that could inadvertently cause the generative model to choose an invalid action. It might reach the box the user wants, but knock books off the shelf in the process.

“We want to allow the user to interact with the robot without introducing those kinds of mistakes, so we get a behavior that is much more aligned with user intent during deployment, but that is also valid and feasible,” Wang says.

Their framework accomplishes this by providing the user with three intuitive ways to correct the robot’s behavior, each of which offers certain advantages.

First, the user can point to the object they want the robot to manipulate in an interface that shows its camera view. Second, they can trace a trajectory in that interface, allowing them to specify how they want the robot to reach the object. Third, they can physically move the robot’s arm in the direction they want it to follow.

“When you are mapping a 2D image of the environment to actions in a 3D space, some information is lost. Physically nudging the robot is the most direct way to specifying user intent without losing any of the information,” says Wang.

Sampling for success

To ensure these interactions don’t cause the robot to choose an invalid action, such as colliding with other objects, the researchers use a specific sampling procedure. This technique lets the model choose an action from the set of valid actions that most closely aligns with the user’s goal.

“Rather than just imposing the user’s will, we give the robot an idea of what the user intends but let the sampling procedure oscillate around its own set of learned behaviors,” Wang explains.

This sampling method enabled the researchers’ framework to outperform the other methods they compared it to during simulations and experiments with a real robot arm in a toy kitchen.

While their method might not always complete the task right away, it offers users the advantage of being able to immediately correct the robot if they see it doing something wrong, rather than waiting for it to finish and then giving it new instructions.

Moreover, after a user nudges the robot a few times until it picks up the correct bowl, it could log that corrective action and incorporate it into its behavior through future training. Then, the next day, the robot could pick up the correct bowl without needing a nudge.

“But the key to that continuous improvement is having a way for the user to interact with the robot, which is what we have shown here,” Wang says.

In the future, the researchers want to boost the speed of the sampling procedure while maintaining or improving its performance. They also want to experiment with robot policy generation in novel environments.

Treating severe or chronic injury to soft tissues such as skin and muscle is a challenge in health care. Current treatment methods can be costly and ineffective, and the frequency of chronic wounds in general from conditions such as diabetes and vascular disease, as well as an increasingly aging population, is only expected to rise.

One promising treatment method involves implanting biocompatible materials seeded with living cells (i.e., microtissue) into the wound. The materials provide a scaffolding for stem cells, or other precursor cells, to grow into the wounded tissue and aid in repair. However, current techniques to construct these scaffolding materials suffer a recurring setback. Human tissue moves and flexes in a unique way that traditional soft materials struggle to replicate, and if the scaffolds stretch, they can also stretch the embedded cells, often causing those cells to die. The dead cells hinder the healing process and can also trigger an inadvertent immune response in the body.

"The human body has this hierarchical structure that actually un-crimps or unfolds, rather than stretches," says Steve Gillmer, a researcher in MIT Lincoln Laboratory's Mechanical Engineering Group. "That's why if you stretch your own skin or muscles, your cells aren't dying. What's actually happening is your tissues are uncrimping a little bit before they stretch."

Gillmer is part of a multidisciplinary research team that is searching for a solution to this stretching setback. He is working with Professor Ming Guo from MIT's Department of Mechanical Engineering and the laboratory's Defense Fabric Discovery Center (DFDC) to knit new kinds of fabrics that can uncrimp and move just as human tissue does.

The idea for the collaboration came while Gillmer and Guo were teaching a course at MIT. Guo had been researching how to grow stem cells on new forms of materials that could mimic the uncrimping of natural tissue. He chose electrospun nanofibers, which worked well, but were difficult to fabricate at long lengths, preventing him from integrating the fibers into larger knit structures for larger-scale tissue repair.

"Steve mentioned that Lincoln Laboratory had access to industrial knitting machines," Guo says. These machines allowed him to switch focus to designing larger knits, rather than individual yarns. "We immediately started to test new ideas through internal support from the laboratory."

Gillmer and Guo worked with the DFDC to discover which knit patterns could move similarly to different types of soft tissue. They started with three basic knit constructions called interlock, rib, and jersey.

"For jersey, think of your T-shirt. When you stretch your shirt, the yarn loops are doing the stretching," says Emily Holtzman, a textile specialist at the DFDC. "The longer the loop length, the more stretch your fabric can accommodate. For ribbed, think of the cuff on your sweater. This fabric construction has a global stretch that allows the fabric to unfold like an accordion."

Interlock is similar to ribbed but is knitted in a denser pattern and contains twice as much yarn per inch of fabric. By having more yarn, there is more surface area on which to embed the cells. "Knit fabrics can also be designed to have specific porosities, or hydraulic permeability, created by the loops of the fabric and yarn sizes," says Erin Doran, another textile specialist on the team. "These pores can help with the healing process as well."

So far, the team has conducted a number of tests embedding mouse embryonic fibroblast cells and mesenchymal stem cells within the different knit patterns and seeing how they behave when the patterns are stretched. Each pattern had variations that affected how much the fabric could uncrimp, in addition to how stiff it became after it started stretching. All showed a high rate of cell survival, and in 2024 the team received an R&D 100 award for their knit designs.

Gillmer explains that although the project began with treating skin and muscle injuries in mind, their fabrics have the potential to mimic many different types of human soft tissue, such as cartilage or fat. The team recently filed a provisional patent that outlines how to create these patterns and identifies the appropriate materials that should be used to make the yarn. This information can be used as a toolbox to tune different knitted structures to match the mechanical properties of the injured tissue to which they are applied.

"This project has definitely been a learning experience for me," Gillmer says. "Each branch of this team has a unique expertise, and I think the project would be impossible without them all working together. Our collaboration as a whole enables us to expand the scope of the work to solve these larger, more complex problems."

A new MIT-led study confirms that the Antarctic ozone layer is healing, as a direct result of global efforts to reduce ozone-depleting substances.

Scientists including the MIT team have observed signs of ozone recovery in the past. But the new study is the first to show, with high statistical confidence, that this recovery is due primarily to the reduction of ozone-depleting substances, versus other influences such as natural weather variability or increased greenhouse gas emissions to the stratosphere.

“There’s been a lot of qualitative evidence showing that the Antarctic ozone hole is getting better. This is really the first study that has quantified confidence in the recovery of the ozone hole,” says study author Susan Solomon, the Lee and Geraldine Martin Professor of Environmental Studies and Chemistry. “The conclusion is, with 95 percent confidence, it is recovering. Which is awesome. And it shows we can actually solve environmental problems.”

The new study appears today in the journal Nature. Graduate student Peidong Wang from the Solomon group in the Department of Earth, Atmospheric and Planetary Sciences (EAPS) is the lead author. His co-authors include Solomon and EAPS Research Scientist Kane Stone, along with collaborators from multiple other institutions.

Roots of ozone recovery

Within the Earth’s stratosphere, ozone is a naturally occurring gas that acts as a sort of sunscreen, protecting the planet from the sun’s harmful ultraviolet radiation. In 1985, scientists discovered a “hole” in the ozone layer over Antarctica that opened up during the austral spring, between September and December. This seasonal ozone depletion was suddenly allowing UV rays to filter down to the surface, leading to skin cancer and other adverse health effects.

In 1986, Solomon, who was then working at the National Oceanic and Atmospheric Administration (NOAA), led expeditions to the Antarctic, where she and her colleagues gathered evidence that quickly confirmed the ozone hole’s cause: chlorofluorocarbons, or CFCs — chemicals that were then used in refrigeration, air conditioning, insulation, and aerosol propellants. When CFCs drift up into the stratosphere, they can break down ozone under certain seasonal conditions.

The following year, those relevations led to the drafting of the Montreal Protocol — an international treaty that aimed to phase out the production of CFCs and other ozone-depleting substances, in hopes of healing the ozone hole.

In 2016, Solomon led a study reporting key signs of ozone recovery. The ozone hole seemed to be shrinking with each year, especially in September, the time of year when it opens up. Still, these observations were qualitative. The study showed large uncertainties regarding how much of this recovery was due to concerted efforts to reduce ozone-depleting substances, or if the shrinking ozone hole was a result of other “forcings,” such as year-to-year weather variability from El Niño, La Niña, and the polar vortex.

“While detecting a statistically significant increase in ozone is relatively straightforward, attributing these changes to specific forcings is more challenging,” says Wang.

Anthropogenic healing

In their new study, the MIT team took a quantitative approach to identify the cause of Antarctic ozone recovery. The researchers borrowed a method from the climate change community, known as “fingerprinting,” which was pioneered by Klaus Hasselmann, who was awarded the Nobel Prize in Physics in 2021 for the technique. In the context of climate, fingerprinting refers to a method that isolates the influence of specific climate factors, apart from natural, meteorological noise. Hasselmann applied fingerprinting to identify, confirm, and quantify the anthropogenic fingerprint of climate change.

Solomon and Wang looked to apply the fingerprinting method to identify another anthropogenic signal: the effect of human reductions in ozone-depleting substances on the recovery of the ozone hole.

“The atmosphere has really chaotic variability within it,” Solomon says. “What we’re trying to detect is the emerging signal of ozone recovery against that kind of variability, which also occurs in the stratosphere.”

The researchers started with simulations of the Earth’s atmosphere and generated multiple “parallel worlds,” or simulations of the same global atmosphere, under different starting conditions. For instance, they ran simulations under conditions that assumed no increase in greenhouse gases or ozone-depleting substances. Under these conditions, any changes in ozone should be the result of natural weather variability. They also ran simulations with only increasing greenhouse gases, as well as only decreasing ozone-depleting substances.

They compared these simulations to observe how ozone in the Antarctic stratosphere changed, both with season, and across different altitudes, in response to different starting conditions. From these simulations, they mapped out the times and altitudes where ozone recovered from month to month, over several decades, and identified a key “fingerprint,” or pattern, of ozone recovery that was specifically due to conditions of declining ozone-depleting substances.

The team then looked for this fingerprint in actual satellite observations of the Antarctic ozone hole from 2005 to the present day. They found that, over time, the fingerprint that they identified in simulations became clearer and clearer in observations. In 2018, the fingerprint was at its strongest, and the team could say with 95 percent confidence that ozone recovery was due mainly to reductions in ozone-depleting substances.

“After 15 years of observational records, we see this signal to noise with 95 percent confidence, suggesting there’s only a very small chance that the observed pattern similarity can be explained by variability noise,” Wang says. “This gives us confidence in the fingerprint. It also gives us confidence that we can solve environmental problems. What we can learn from ozone studies is how different countries can swiftly follow these treaties to decrease emissions.”

If the trend continues, and the fingerprint of ozone recovery grows stronger, Solomon anticipates that soon there will be a year, here and there, when the ozone layer stays entirely intact. And eventually, the ozone hole should stay shut for good.

“By something like 2035, we might see a year when there’s no ozone hole depletion at all in the Antarctic. And that will be very exciting for me,” she says. “And some of you will see the ozone hole go away completely in your lifetimes. And people did that.”

This research was supported, in part, by the National Science Foundation and NASA.

Building on more than two decades of research, a study by MIT neuroscientists at The Picower Institute for Learning and Memory reports a new way to treat pathology and symptoms of fragile X syndrome, the most common genetically-caused autism spectrum disorder. The team showed that augmenting a novel type of neurotransmitter signaling reduced hallmarks of fragile X in mouse models of the disorder.

The new approach, described in Cell Reports, works by targeting a specific molecular subunit of “NMDA” receptors that they discovered plays a key role in how neurons synthesize proteins to regulate their connections, or “synapses,” with other neurons in brain circuits. The scientists showed that in fragile X model mice, increasing the receptor’s activity caused neurons in the hippocampus region of the brain to increase molecular signaling that suppressed excessive bulk protein synthesis, leading to other key improvements.

Setting the table

“One of the things I find most satisfying about this study is that the pieces of the puzzle fit so nicely into what had come before,” says study senior author Mark Bear, Picower Professor in MIT’s Department of Brain and Cognitive Sciences. Former postdoc Stephanie Barnes, now a lecturer at the University of Glasgow, is the study’s lead author.

Bear’s lab studies how neurons continually edit their circuit connections, a process called “synaptic plasticity” that scientists believe to underlie the brain’s ability to adapt to experience and to form and process memories. These studies led to two discoveries that set the table for the newly published advance. In 2011, Bear’s lab showed that fragile X and another autism disorder, tuberous sclerosis (Tsc), represented two ends of a continuum of a kind of protein synthesis in the same neurons. In fragile X there was too much. In Tsc there was too little. When lab members crossbred fragile X and Tsc mice, in fact, their offspring emerged healthy, as the mutations of each disorder essentially canceled each other out.

More recently, Bear’s lab showed a different dichotomy. It has long been understood from their influential work in the 1990s that the flow of calcium ions through NMDA receptors can trigger a form of synaptic plasticity called “long-term depression” (LTD). But in 2020, they found that another mode of signaling by the receptor — one that did not require ion flow — altered protein synthesis in the neuron and caused a physical shrinking of the dendritic “spine” structures housing synapses.

For Bear and Barnes, these studies raised the prospect that if they could pinpoint how NMDA receptors affect protein synthesis they might identify a new mechanism that could be manipulated therapeutically to address fragile X (and perhaps tuberous sclerosis) pathology and symptoms. That would be an important advance to complement ongoing work Bear’s lab has done to correct fragile X protein synthesis levels via another receptor called mGluR5.

Receptor dissection

In the new study, Bear and Barnes’ team decided to use the non-ionic effect on spine shrinkage as a readout to dissect how NMDARs signal protein synthesis for synaptic plasticity in hippocampus neurons. They hypothesized that the dichotomy of ionic effects on synaptic function and non-ionic effects on spine structure might derive from the presence of two distinct components of NMDA receptors: “subunits” called GluN2A and GluN2B. To test that, they used genetic manipulations to knock out each of the subunits. When they did so, they found that knocking out “2A” or “2B” could eliminate LTD, but that only knocking out 2B affected spine size. Further experiments clarified that 2A and 2B are required for LTD, but that spine shrinkage solely depends on the 2B subunit.

The next task was to resolve how the 2B subunit signals spine shrinkage. A promising possibility was a part of the subunit called the “carboxyterminal domain,” or CTD. So, in a new experiment Bear and Barnes took advantage of a mouse that had been genetically engineered by researchers at the University of Edinburgh so that the 2A and 2B CTDs could be swapped with one another. A telling result was that when the 2B subunit lacked its proper CTD, the effect on spine structure disappeared. The result affirmed that the 2B subunit signals spine shrinkage via its CTD.

Another consequence of replacing the CTD of the 2B subunit was an increase in bulk protein synthesis that resembled findings in fragile X. Conversely, augmenting the non-ionic signaling through the 2B subunit suppressed bulk protein synthesis, reminiscent of Tsc.

Treating fragile X

Putting the pieces together, the findings indicated that augmenting signaling through the 2B subunit might, like introducing the mutation causing Tsc, rescue aspects of fragile X.

Indeed, when the scientists swapped in the 2B subunit CTD of NMDA receptor in fragile X model mice they found correction of not only the excessive bulk protein synthesis, but also altered synaptic plasticity, and increased electrical excitability that are hallmarks of the disease. To see if a treatment that targets NMDA receptors might be effective in fragile X, they tried an experimental drug called Glyx-13. This drug binds to the 2B subunit of NMDA receptors to augment signaling. The researchers found that this treatment can also normalize protein synthesis and reduced sound-induced seizures in the fragile X mice.

The team now hypothesizes, based on another prior study in the lab, that the beneficial effect to fragile X mice of the 2B subunit’s CTD signaling is that it shifts the balance of protein synthesis away from an all-too-efficient translation of short messenger RNAs (which leads to excessive bulk protein synthesis) toward a lower-efficiency translation of longer messenger RNAs.

Bear says he does not know what the prospects are for Glyx-13 as a clinical drug, but he noted that there are some drugs in clinical development that specifically target the 2B subunit of NMDA receptors.

In addition to Bear and Barnes, the study’s other authors are Aurore Thomazeau, Peter Finnie, Max Heinreich, Arnold Heynen, Noboru Komiyama, Seth Grant, Frank Menniti, and Emily Osterweil.

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

For as long as people have been communicating through writing, they have found ways to keep their messages private. Before the invention of the gummed envelope in 1830, securing correspondence involved letterlocking, an ingenious process of folding a flat sheet of paper to become its own envelope, often using a combination of folds, tucks, slits, or adhesives such as sealing wax. Letter writers from Erasmus to Catherine de’ Medici to Emily Dickinson employed these techniques, which Jana Dambrogio, the MIT Libraries’ Thomas F. Peterson (1957) Conservator, has named “letterlocking.”

“The study of letterlocking very consciously bridges humanities and sciences,” says Dambrogio, who first became interested in the practice as a fellow in the conservation studio of the Vatican Apostolic Archives, where she discovered examples from the 15th and 16th centuries. “It draws on the perspectives of not only conservators and historians, but also engineers, imaging experts, and scientists.”

Now the rich history of this centuries-old document security technology is the subject of a new book, “Letterlocking: The Hidden History of the Letter,” published by the MIT Press and co-authored with Daniel Starza Smith, a lecturer in early modern English literature at King’s College London. Dambrogio and Smith have pioneered the field of letterlocking research over the last 10 years, working with an international and interdisciplinary collection of experts, the Unlocking History Research Group.

With more than 300 images and diagrams, “Letterlocking” explores the practice’s history through real examples from all over the world. It includes a dictionary of 60 technical terms and concepts, systems the authors developed while studying more than 250,000 historic letters. The book aims to be a springboard for new discoveries, whether providing a new lens on history or spurring technological advancements.

In working with the Brienne Collection — a 17th-century postal trunk full of undelivered letters — the Unlocking History Research Group sought to study intact examples of locked letters without destroying them in the process. This stimulated advances in conservation, radiology, and computational algorithms. In 2020, the team collaborated with Amanda Ghassaei SM ’17 and Holly Jackson ’22, working at the MIT Center for Bits and Atoms, and students and faculty from the MIT Computer Science and Artificial Intelligence Laboratory; the School of Humanities, Arts, and Social Sciences; and the Department of Materials Science and Engineering to develop new algorithms that could virtually read an unopened letter, publishing the results in Nature Communications in 2021.

“Letterlocking” also offers a comprehensive guide to making one’s own locked letters. “The best introduction to letterlocking is to make some models,” says Dambrogio. “Feel the shape and the weight; see how easy it would be to conceal or hard to open without being noticed. We’re inviting people to explore and expand this new field of study through ‘mind and hand.’”

When Louis DeRidder was 12 years old, he had a medical emergency that nearly cost him his life. The terrifying experience gave him a close-up look at medical care and made him eager to learn more.

“You can’t always pinpoint exactly what gets you interested in something, but that was a transformative moment,” says DeRidder.

In high school, he grabbed the chance to participate in a medicine-focused program, spending about half of his days during his senior year in high school learning about medical science and shadowing doctors.

DeRidder was hooked. He became fascinated by the technologies that make treatments possible and was particularly interested in how drugs are delivered to the brain, a curiosity that sparked a lifelong passion.

“Here I was, a 17-year-old in high school, and a decade later, that problem still fascinates me,” he says. “That’s what eventually got me into the drug delivery field.”

DeRidder’s interests led him to transfer half-way through his undergraduate studies to Johns Hopkins University, where he performed research he had proposed in a Goldwater Scholarship proposal. The research focused on the development of a nanoparticle-drug conjugate to deliver a drug to brain cells in order to transform them from a pro-inflammatory to an anti-inflammatory phenotype. Such a technology could be valuable in the treatment of neurodegenerative diseases, including Alzheimer’s and Parkinson’s.

In 2019, DeRidder entered the joint Harvard-MIT Health Sciences and Technology program, where he has embarked on a somewhat different type of drug delivery project — developing a device that measures the concentration of a chemotherapy drug in the blood while it is being administered and adjusts the infusion rate so the concentration is optimal for the patient. The system is known as CLAUDIA, or Closed-Loop AUtomated Drug Infusion RegulAtor, and can allow for the personalization of drug dosing for a variety of different drugs.

The project stemmed from discussions with his faculty advisors — Robert Langer, the David H. Koch Institute Professor, and Giovanni Traverso, the Karl Van Tassel Career Development Professor and a gastroenterologist at Brigham and Women’s Hospital. They explained to him that chemotherapy dosing is based on a formula developed in 1916 that estimates a patient’s body surface area. The formula doesn’t consider important influences such as differences in body composition and metabolism, or circadian fluctuations that can affect how a drug interacts with a patient.

“Once my advisors presented the reality of how chemotherapies are dosed,” DeRidder says, “I thought, ‘This is insane. How is this the clinical reality?’”

He and his advisors agreed this was a great project for his PhD.

“After they gave me the problem statement, we began to brainstorm ways that we could develop a medical device to improve the lives of patients” DeRidder says, adding, “I love starting with a blank piece of paper and then brainstorming to work out the best solution.”

Almost from the start, DeRidder’s research process involved MATLAB and Simulink, developed by the mathematical computer software company MathWorks.

“MathWorks and Simulink are key to what we do,” DeRidder says. “They enable us to model the drug pharmacokinetics — how the body distributes and metabolizes the drug. We also model the components of our system with their software. That was especially critical for us in the very early days, because it let us know whether it was even possible to control the concentration of the drug. And since then, we’ve continuously improved the control algorithm, using these simulations. You simulate hundreds of different experiments before performing any experiments in the lab.”

With his innovative use of the MATLAB and Simulink tools, DeRidder was awarded MathWorks fellowships both last year and this year. He has also received a National Science Foundation Graduate Research Fellowship.

“The fellowships have been critical to our development of the CLAUDIA drug-delivery system,” DeRidder says, adding that he has “had the pleasure of working with a great team of students and researchers in the lab.”

He says he would like to move CLAUDIA toward clinical use, where he thinks it could have significant impact. “Whatever I can do to help push it toward the clinic, including potentially helping to start a company to help commercialize the system, I’m definitely interested in doing it.”

In addition to developing CLAUDIA, DeRidder is working on developing new nanoparticles to deliver therapeutic nucleic acids. The project involves synthesizing new nucleic acid molecules, as well as developing the new polymeric and lipid nanoparticles to deliver the nucleic acids to targeted tissue and cells.

DeRidder says he likes working on technologies at different scales, from medical devices to molecules — all with the potential to improve the practice of medicine.

Meanwhile, he finds time in his busy schedule to do community service. For the past three years, he has spent time helping the homeless on Boston streets.

“It’s easy to lose track of the concrete, simple ways that we can serve our communities when we’re doing research,” DeRidder says, “which is why I have often sought out ways to serve people I come across every day, whether it is a student I mentor in lab, serving the homeless, or helping out the stranger you meet in the store who is having a bad day.”

Ultimately, DeRidder says, he’ll head back to work that also recalls his early exposure to the medical field in high school, where he interacted with a lot of people with different types of dementia and other neurological diseases at a local nursing home.

“My long-term plan includes working on developing devices and molecular therapies to treat neurological diseases, in addition to continuing to work on cancer,” he says. “Really, I’d say that early experience had a big impact on me.”

In biology, seeing can lead to understanding, and researchers in Professor Edward Boyden’s lab at the McGovern Institute for Brain Research are committed to bringing life into sharper focus. With a pair of new methods, they are expanding the capabilities of expansion microscopy — a high-resolution imaging technique the group introduced in 2015 — so researchers everywhere can see more when they look at cells and tissues under a light microscope.

“We want to see everything, so we’re always trying to improve it,” says Boyden, the Y. Eva Tan Professor in Neurotechnology at MIT. “A snapshot of all life, down to its fundamental building blocks, is really the goal.” Boyden is also a Howard Hughes Medical Institute investigator and a member of the Yang Tan Collective at MIT. 

With new ways of staining their samples and processing images, users of expansion microscopy can now see vivid outlines of the shapes of cells in their images and pinpoint the locations of many different proteins inside a single tissue sample with resolution that far exceeds that of conventional light microscopy. These advances, both reported in open-access form in the journal Nature Communications, enable new ways of tracing the slender projections of neurons and visualizing spatial relationships between molecules that contribute to health and disease.

Expansion microscopy uses a water-absorbing hydrogel to physically expand biological tissues. After a tissue sample has been permeated by the hydrogel, it is hydrated. The hydrogel swells as it absorbs water, preserving the relative locations of molecules in the tissue as it gently pulls them away from one another. As a result, crowded cellular components appear separate and distinct when the expanded tissue is viewed under a light microscope. The approach, which can be performed using standard laboratory equipment, has made super-resolution imaging accessible to most research teams.

Since first developing expansion microscopy, Boyden and his team have continued to enhance the method — increasing its resolution, simplifying the procedure, devising new features, and integrating it with other tools.

Visualizing cell membranes

One of the team’s latest advances is a method called ultrastructural membrane expansion microscopy (umExM), which they described in the Feb. 12 issue of Nature Communications. With it, biologists can use expansion microscopy to visualize the thin membranes that form the boundaries of cells and enclose the organelles inside them. These membranes, built mostly of molecules called lipids, have been notoriously difficult to densely label in intact tissues for imaging with light microscopy. Now, researchers can use umExM to study cellular ultrastructure and organization within tissues.

Tay Shin SM ’20, PhD ’23, a former graduate student in Boyden’s lab and a J. Douglas Tan Fellow in the Tan-Yang Center for Autism Research at MIT, led the development of umExM. “Our goal was very simple at first: Let’s label membranes in intact tissue, much like how an electron microscope uses osmium tetroxide to label membranes to visualize the membranes in tissue,” he says. “It turns out that it’s extremely hard to achieve this.”

The team first needed to design a label that would make the membranes in tissue samples visible under a light microscope. “We almost had to start from scratch,” Shin says. “We really had to think about the fundamental characteristics of the probe that is going to label the plasma membrane, and then think about how to incorporate them into expansion microscopy.” That meant engineering a molecule that would associate with the lipids that make up the membrane and link it to both the hydrogel used to expand the tissue sample and a fluorescent molecule for visibility.

After optimizing the expansion microscopy protocol for membrane visualization and extensively testing and improving potential probes, Shin found success one late night in the lab. He placed an expanded tissue sample on a microscope and saw sharp outlines of cells.

Because of the high resolution enabled by expansion, the method allowed Boyden’s team to identify even the tiny dendrites that protrude from neurons and clearly see the long extensions of their slender axons. That kind of clarity could help researchers follow individual neurons’ paths within the densely interconnected networks of the brain, the researchers say.

Boyden calls tracing these neural processes “a top priority of our time in brain science.” Such tracing has traditionally relied heavily on electron microscopy, which requires specialized skills and expensive equipment. Shin says that because expansion microscopy uses a standard light microscope, it is far more accessible to laboratories worldwide.

Shin and Boyden point out that users of expansion microscopy can learn even more about their samples when they pair the new ability to reveal lipid membranes with fluorescent labels that show where specific proteins are located. “That’s important, because proteins do a lot of the work of the cell, but you want to know where they are with respect to the cell’s structure,” Boyden says.

One sample, many proteins

To that end, researchers no longer have to choose just a few proteins to see when they use expansion microscopy. With a new method called multiplexed expansion revealing (multiExR), users can now label and see more than 20 different proteins in a single sample. Biologists can use the method to visualize sets of proteins, see how they are organized with respect to one another, and generate new hypotheses about how they might interact.

A key to that new method, reported Nov. 9, 2024, in Nature Communications, is the ability to repeatedly link fluorescently labeled antibodies to specific proteins in an expanded tissue sample, image them, then strip these away and use a new set of antibodies to reveal a new set of proteins. Postdoc Jinyoung Kang fine-tuned each step of this process, assuring tissue samples stayed intact and the labeled proteins produced bright signals in each round of imaging.

After capturing many images of a single sample, Boyden’s team faced another challenge: how to ensure those images were in perfect alignment so they could be overlaid with one another, producing a final picture that showed the precise positions of all of the proteins that had been labeled and visualized one by one.

Expansion microscopy lets biologists visualize some of cells’ tiniest features — but to find the same features over and over again during multiple rounds of imaging, Boyden’s team first needed to home in on a larger structure. “These fields of view are really tiny, and you’re trying to find this really tiny field of view in a gel that’s actually become quite large once you’ve expanded it,” explains Margaret Schroeder, a graduate student in Boyden’s lab who, with Kang, led the development of multiExR.

To navigate to the right spot every time, the team decided to label the blood vessels that pass through each tissue sample and use these as a guide. To enable precise alignment, certain fine details also needed to consistently appear in every image; for this, the team labeled several structural proteins. With these reference points and customized imaging processing software, the team was able to integrate all of their images of a sample into one, revealing how proteins that had been visualized separately were arranged relative to one another.

The team used multiExR to look at amyloid plaques — the aberrant protein clusters that notoriously develop in brains affected by Alzheimer’s disease. “We could look inside those amyloid plaques and ask, what’s inside of them? And because we can stain for many different proteins, we could do a high-throughput exploration,” Boyden says. The team chose 23 different proteins to view in their images. The approach revealed some surprises, such as the presence of certain neurotransmitter receptors (AMPARs). “Here’s one of the most famous receptors in all of neuroscience, and there it is, hiding out in one of the most famous molecular hallmarks of pathology in neuroscience,” says Boyden. It’s unclear what role, if any, the receptors play in Alzheimer’s disease — but the finding illustrates how the ability to see more inside cells can expose unexpected aspects of biology and raise new questions for research.

Funding for this work came from MIT, Lisa Yang and Y. Eva Tan, John Doerr, the Open Philanthropy Project, the Howard Hughes Medical Institute, the U.S. Army, Cancer Research U.K., the New York Stem Cell Foundation, the U.S. National Institutes of Health, Lore McGovern, Good Ventures, Schmidt Futures, Samsung, MathWorks, the Collamore-Rogers Fellowship, the U.S. National Science Foundation, Alana Foundation USA, the Halis Family Foundation, Lester A. Gimpelson, Donald and Glenda Mattes, David B. Emmes, Thomas A. Stocky, Avni U. Shah, Kathleen Octavio, Good Ventures/Open Philanthropy, and the European Union’s Horizon 2020 program.

The following is a joint announcement from the MIT Microsystems Technology Laboratories and GlobalFoundries. 

MIT and GlobalFoundries (GF), a leading manufacturer of essential semiconductors, have announced a new research agreement to jointly pursue advancements and innovations for enhancing the performance and efficiency of critical semiconductor technologies. The collaboration will be led by MIT’s Microsystems Technology Laboratories (MTL) and GF’s research and development team, GF Labs.

With an initial research focus on artificial intelligence and other applications, the first projects are expected to leverage GF’s differentiated silicon photonics technology, which monolithically integrates radio frequency silicon-on-insulator (RF SOI), CMOS (complementary metal-oxide semiconductor), and optical features on a single chip to realize power efficiencies for data centers, and GF’s 22FDX platform, which delivers ultra-low power consumption for intelligent devices at the edge.

“The collaboration between MIT MTL and GF exemplifies the power of academia-industry cooperation in tackling the most pressing challenges in semiconductor research,” says Tomás Palacios, MTL director and the Clarence J. LeBel Professor of Electrical Engineering and Computer Science. Palacios will serve as the MIT faculty lead for this research initiative.

“By bringing together MIT's world-renowned capabilities with GF's leading semiconductor platforms, we are positioned to drive significant research advancements in GF’s essential chip technologies for AI,” says Gregg Bartlett, chief technology officer at GF. “This collaboration underscores our commitment to innovation and highlights our dedication to developing the next generation of talent in the semiconductor industry. Together, we will research transformative solutions in the industry.”

“Integrated circuit technologies are the core driving a broad spectrum of applications ranging from mobile computing and communication devices to automotive, energy, and cloud computing,” says Anantha P. Chandrakasan, dean of MIT's School of Engineering, chief innovation and strategy officer, and the Vannevar Bush Professor of Electrical Engineering and Computer Science. “This collaboration allows MIT’s exceptional research community to leverage GlobalFoundries’ wide range of industry domain experts and advanced process technologies to drive exciting innovations in microelectronics across domains — while preparing our students to take on leading roles in the workforce of the future.”

The new research agreement was formalized at a signing ceremony on campus at MIT. It builds upon GF’s successful past and ongoing engagements with the university. GF serves on MTL’s Microsystems Industrial Group, which brings together industry and academia to engage in research. MIT faculty are active participants in GF’s University Partnership Program focused on joint semiconductor research and prototyping. Additionally, GF and MIT collaborate on several workforce development initiatives, including through the Northeast Microelectronics Coalition, a U.S. Department of Defense Microelectronics Commons Hub.

High-temperature superconducting magnets made from REBCO, an acronym for rare earth barium copper oxide, make it possible to create an intense magnetic field that can confine the extremely hot plasma needed for fusion reactions, which combine two hydrogen atoms to form an atom of helium, releasing a neutron in the process.

But some early tests suggested that neutron irradiation inside a fusion power plant might instantaneously suppress the superconducting magnets’ ability to carry current without resistance (called critical current), potentially causing a reduction in the fusion power output.

Now, a series of experiments has clearly demonstrated that this instantaneous effect of neutron bombardment, known as the “beam on effect,” should not be an issue during reactor operation, thus clearing the path for projects such as the ARC fusion system being developed by MIT spinoff company Commonwealth Fusion Systems.

The findings were reported in the journal Superconducting Science and Technology, in a paper by MIT graduate student Alexis Devitre and professors Michael Short, Dennis Whyte, and Zachary Hartwig, along with six others.

“Nobody really knew if it would be a concern,” Short explains. He recalls looking at these early findings: “Our group thought, man, somebody should really look into this. But now, luckily, the result of the paper is: It’s conclusively not a concern.”

The possible issue first arose during some initial tests of the REBCO tapes planned for use in the ARC system. “I can remember the night when we first tried the experiment,” Devitre recalls. “We were all down in the accelerator lab, in the basement. It was a big shocker because suddenly the measurement we were looking at, the critical current, just went down by 30 percent” when it was measured under radiation conditions (approximating those of the fusion system), as opposed to when it was only measured after irradiation.

Before that, researchers had irradiated the REBCO tapes and then tested them afterward, Short says. “We had the idea to measure while irradiating, the way it would be when the reactor’s really on,” he says. “And then we observed this giant difference, and we thought, oh, this is a big deal. It’s a margin you’d want to know about if you’re designing a reactor.”

After a series of carefully calibrated tests, it turned out the drop in critical current was not caused by the irradiation at all, but was just an effect of temperature changes brought on by the proton beam used for the irradiation experiments. This is something that would not be a factor in an actual fusion plant, Short says.

“We repeated experiments ‘oh so many times’ and collected about a thousand data points,” Devitre says. They then went through a detailed statistical analysis to show that the effects were exactly the same, under conditions where the material was just heated as when it was both heated and irradiated.

This excluded the possibility that the instantaneous suppression of the critical current had anything to do with the “beam on effect,” at least within the sensitivity of their tests. “Our experiments are quite sensitive,” Short says. “We can never say there’s no effect, but we can say that there’s no important effect.”

To carry out these tests required building a special facility for the purpose. Only a few such facilities exist in the world. “They’re all custom builds, and without this, we wouldn’t have been able to find out the answer,” he says.

The finding that this specific issue is not a concern for the design of fusion plants “illustrates the power of negative results. If you can conclusively prove that something doesn’t happen, you can stop scientists from wasting their time hunting for something that doesn’t exist.” And in this case, Short says, “You can tell the fusion companies: ‘You might have thought this effect would be real, but we’ve proven that it’s not, and you can ignore it in your designs.’ So that’s one more risk retired.”

That could be a relief to not only Commonwealth Fusion Systems but also several other companies that are also pursuing fusion plant designs, Devitre says. “There’s a bunch. And it’s not just fusion companies,” he adds. There remains the important issue of longer-term degradation of the REBCO that would occur over years or decades, which the group is presently investigating. Others are pursuing the use of these magnets for satellite thrusters and particle accelerators to study subatomic physics, where the effect could also have been a concern. For all these uses, “this is now one less thing to be concerned about,” Devitre says.

The research team also included David Fischer, Kevin Woller, Maxwell Rae, Lauryn Kortman, and Zoe Fisher at MIT, and N. Riva at Proxima Fusion in Germany. This research was supported by Eni S.p.A. through the MIT Energy Initiative.

A vast search of natural diversity has led scientists at MIT’s McGovern Institute for Brain Research and the Broad Institute of MIT and Harvard to uncover ancient systems with potential to expand the genome editing toolbox. 

These systems, which the researchers call TIGR (Tandem Interspaced Guide RNA) systems, use RNA to guide them to specific sites on DNA. TIGR systems can be reprogrammed to target any DNA sequence of interest, and they have distinct functional modules that can act on the targeted DNA. In addition to its modularity, TIGR is very compact compared to other RNA-guided systems, like CRISPR, which is a major advantage for delivering it in a therapeutic context.  

These findings are reported online Feb. 27 in the journal Science.

“This is a very versatile RNA-guided system with a lot of diverse functionalities,” says Feng Zhang, the James and Patricia Poitras Professor of Neuroscience at MIT, who led the research. The TIGR-associated (Tas) proteins that Zhang’s team found share a characteristic RNA-binding component that interacts with an RNA guide that directs it to a specific site in the genome. Some cut the DNA at that site, using an adjacent DNA-cutting segment of the protein. That modularity could facilitate tool development, allowing researchers to swap useful new features into natural Tas proteins.

“Nature is pretty incredible,” says Zhang, who is also an investigator at the McGovern Institute and the Howard Hughes Medical Institute, a core member of the Broad Institute, a professor of brain and cognitive sciences and biological engineering at MIT, and co-director of the K. Lisa Yang and Hock E. Tan Center for Molecular Therapeutics at MIT. “It’s got a tremendous amount of diversity, and we have been exploring that natural diversity to find new biological mechanisms and harnessing them for different applications to manipulate biological processes,” he says. Previously, Zhang’s team adapted bacterial CRISPR systems into gene editing tools that have transformed modern biology. His team has also found a variety of programmable proteins, both from CRISPR systems and beyond. 

In their new work, to find novel programmable systems, the team began by zeroing in a structural feature of the CRISPR-Cas9 protein that binds to the enzyme’s RNA guide. That is a key feature that has made Cas9 such a powerful tool: “Being RNA-guided makes it relatively easy to reprogram, because we know how RNA binds to other DNA or other RNA,” Zhang explains. His team searched hundreds of millions of biological proteins with known or predicted structures, looking for any that shared a similar domain. To find more distantly related proteins, they used an iterative process: from Cas9, they identified a protein called IS110, which had previously been shown by others to bind RNA. They then zeroed in on the structural features of IS110 that enable RNA binding and repeated their search. 

At this point, the search had turned up so many distantly related proteins that they team turned to artificial intelligence to make sense of the list. “When you are doing iterative, deep mining, the resulting hits can be so diverse that they are difficult to analyze using standard phylogenetic methods, which rely on conserved sequence,” explains Guilhem Faure, a computational biologist in Zhang’s lab. With a protein large language model, the team was able to cluster the proteins they had found into groups according to their likely evolutionary relationships. One group set apart from the rest, and its members were particularly intriguing because they were encoded by genes with regularly spaced repetitive sequences reminiscent of an essential component of CRISPR systems. These were the TIGR-Tas systems.

Zhang’s team discovered more than 20,000 different Tas proteins, mostly occurring in bacteria-infecting viruses. Sequences within each gene’s repetitive region — its TIGR arrays — encode an RNA guide that interacts with the RNA-binding part of the protein. In some, the RNA-binding region is adjacent to a DNA-cutting part of the protein. Others appear to bind to other proteins, which suggests they might help direct those proteins to DNA targets.     

Zhang and his team experimented with dozens of Tas proteins, demonstrating that some can be programmed to make targeted cuts to DNA in human cells. As they think about developing TIGR-Tas systems into programmable tools, the researchers are encouraged by features that could make those tools particularly flexible and precise.

They note that CRISPR systems can only be directed to segments of DNA that are flanked by short motifs known as PAMs (protospacer adjacent motifs). TIGR Tas proteins, in contrast, have no such requirement. “This means theoretically, any site in the genome should be targetable,” says scientific advisor Rhiannon Macrae. The team’s experiments also show that TIGR systems have what Faure calls a “dual-guide system,” interacting with both strands of the DNA double helix to home in on their target sequences, which should ensure they act only where they are directed by their RNA guide. What’s more, Tas proteins are compact — a quarter of the size Cas9, on average — making them easier to deliver, which could overcome a major obstacle to therapeutic deployment of gene editing tools.  

Excited by their discovery, Zhang’s team is now investigating the natural role of TIGR systems in viruses, as well as how they can be adapted for research or therapeutics. They have determined the molecular structure of one of the Tas proteins they found to work in human cells, and will use that information to guide their efforts to make it more efficient. Additionally, they note connections between TIGR-Tas systems and certain RNA-processing proteins in human cells. “I think there’s more there to study in terms of what some of those relationships may be, and it may help us better understand how these systems are used in humans,” Zhang says.

This work was supported by the Helen Hay Whitney Foundation, Howard Hughes Medical Institute, K. Lisa Yang and Hock E. Tan Center for Molecular Therapeutics, Broad Institute Programmable Therapeutics Gift Donors, Pershing Square Foundation, William Ackman, Neri Oxman, the Phillips family, J. and P. Poitras, and the BT Charitable Foundation. 

One large metropolis might have several different train systems, from local intercity lines to commuter trains to longer regional lines.

When designing a system of train tracks, stations, and schedules in this network, should rail operators assume each entity operates independently, seeking only to maximize its own revenue? Or that they fully cooperate all the time with a joint plan, putting their own interest aside?

In the real world, neither assumption is very realistic.

Researchers from MIT and ETH Zurich have developed a new planning tool that mixes competition and cooperation to help operators in a complex, multiregional network strategically determine when and how they should work together.

Their framework is unusual because it incorporates co-investment and payoff-sharing mechanisms that identify which joint infrastructure projects a stakeholder should invest in with other operators to maximize collective benefits. The tool can help mobility stakeholders, such as governments, transport agencies, and firms, determine the right time to collaborate, how much they should invest in cooperative projects, how the profits should be distributed, and what would happen if they withdrew from the negotiations.

“It might seem counterintuitive, but sometimes you want to invest in your opponent so that, at some point, this investment will come back to you. Thanks to game theory, one can formalize this intuition to give rise to an interesting class of problems,” says Gioele Zardini, the Rudge and Nancy Allen Assistant Professor of Civil and Environmental Engineering at MIT, a principal investigator in the Laboratory for Information and Decision Systems (LIDS), an affiliate faculty with the Institute for Data, Systems, and Society (IDSS), and senior author of a paper on this planning framework.

Numerical analysis shows that, by investing a portion of their budget into some shared infrastructure projects, independent operators can earn more revenue than if they operated completely noncooperatively.

In the example of the rail operators, the researchers demonstrate that co-investment also benefits users by improving regional train service. This win-win situation encourages more people to take the train, boosting revenues for operators and reducing emissions from automobiles, says Mingjia He, a graduate student at ETH Zurich and lead author.

“The key point here is that transport network design is not a zero-sum game. One operator’s gain doesn’t have to mean the others’ loss. By shifting the perception from isolated, self-optimization to strategic interaction, cooperation can create greater value for everyone involved,” she says.

Beyond transportation, this planning framework could help companies in a crowded industry or governments of neighboring countries test co-investment strategies.

He and Zardini are joined on the paper by ETH Zurich researchers Andrea Censi and Emilio Frazzoli. The research will be presented at the 2025 American Control Conference (ACC), and the paper has been selected as a Student Best Paper Award finalist.

Mixing cooperation and competition

Building transportation infrastructure in a multiregional network typically requires a huge investment of time and resources. Major infrastructure projects have an outsized impact that can stretch far beyond one region or operator.

Each region has its own priorities and decision-makers, such as local transportation authorities, which often results in the failure of coordination.

“If local systems are designed separately, regional travel may be more difficult, making the whole system less efficient. But if self-interested stakeholders don’t benefit from coordination, they are less likely to support the plan,” He says.

To find the best mix of cooperation and competition, the researchers used game theory to build a framework that enables operators to align interests and improve regional cooperation in a way that benefits all.

For instance, last year the Swiss government agreed to invest 50 million euros to electrify and expand part of a regional rail network in Germany, with the goal of creating a faster rail connection between three Swiss cities.

The researchers’ planning framework could help independent entities, from regional governments to rail operators, identify when and how to undertake such collaborations.

The first step involves simulating the outcomes if operators don’t collaborate. Then, using the co-investment and payoff-sharing mechanisms, the decision-maker can explore cooperative approaches.

To identify a fair way to split revenues from shared projects, the researchers design a payoff-sharing mechanism based on a game theory concept known as the Nash bargaining solution. This technique will determine how much benefit operators would receive in different cooperative scenarios, taking into account the benefits they would achieve with no collaboration.

The benefits of co-investment

Once they had designed the planning framework, the researchers tested it on a simulated transportation network with multiple competing rail operators. They assessed various co-investment ratios across multiple years to identify the best decisions for operators.

In the end, they found that a semicooperative approach leads to the highest returns for all stakeholders. For instance, in one scenario, by co-investing 50 percent of their total budgets into shared infrastructure projects, all operators maximized their returns.

In another scenario, they show that by investing just 3.3 percent of their total budget in the first year of a multiyear cooperative project, operators can boost outcomes by 30 percent across three metrics: revenue, reduced costs for customers, and lower emissions.

“This proves that a small, up-front investment can lead to significant long-term benefits,” He says.

When they applied their framework to more realistic multiregional networks where all regions weren’t the same size, this semicooperative approach achieved even better results.

However, their analyses indicate that returns don’t increase in a linear way — sometimes increasing the co-investment ratio does not increase the benefit for operators.

Success is a multifaceted issue that depends on how much is invested by all operators, which projects are chosen, when investment happens, and how the budget is distributed over time, He explains.

“These strategic decisions are complex, which is why simulations and optimization are necessary to find the best cooperation and negotiation strategies. Our framework can help operators make smarter investment choices and guide them through the negotiation process,” she says.

The framework could also be applied to other complex network design problems, such as in communications or energy distribution.

In the future, the researchers want to build a user-friendly interface that will allow a stakeholder to easily explore different collaborative options. They also want to consider more complex scenarios, such as the role policy plays in shared infrastructure decisions or the robust cooperative strategies that handle risks and uncertainty.

This work was supported, in part, by the ETH Zurich Mobility Initiative and the ETH Zurich Foundation.

MIT physicists report the unexpected discovery of electrons forming crystalline structures in a material only billionths of a meter thick. The work adds to a gold mine of discoveries originating from the material, which the same team discovered about three years ago.

In a paper published Jan. 22 in Nature, the team describes how electrons in devices made, in part, of the material can become solid, or form crystals, by changing the voltage applied to the devices when they are kept at a temperature similar to that of outer space. Under the same conditions, they also showed the emergence of two new electronic states that add to work they reported last year showing that electrons can split into fractions of themselves.

The physicists were able to make the discoveries thanks to new custom-made filters for better insulation of the equipment involved in the work. These allowed them to cool their devices to a temperature an order of magnitude colder than they achieved for the earlier results.

The team also observed all of these phenomena using two slightly different “versions” of the material, one composed of five layers of atomically thin carbon; the other composed of four layers. This indicates “that there’s a family of materials where you can get this kind of behavior, which is exciting,” says Long Ju, an assistant professor in the MIT Department of Physics who led the work. Ju is also affiliated with MIT’s Materials Research Laboratory and Research Lab of Electronics.

Referring to the material, known as rhombohedral pentalayer graphene, Ju says, “We found a gold mine, and every scoop is revealing something new.”

New material

Rhombohedral pentalayer graphene is essentially a special form of pencil lead. Pencil lead, or graphite, is composed of graphene, a single layer of carbon atoms arranged in hexagons resembling a honeycomb structure. Rhombohedral pentalayer graphene is composed of five layers of graphene stacked in a specific overlapping order.

Since Ju and colleagues discovered the material, they have tinkered with it by adding layers of another material they thought might accentuate the graphene’s properties, or even produce new phenomena. For example, in 2023 they created a sandwich of rhombohedral pentalayer graphene with “buns” made of hexagonal boron nitride. By applying different voltages, or amounts of electricity, to the sandwich, they discovered three important properties never before seen in natural graphite.

Last year, Ju and colleagues reported yet another important and even more surprising phenomenon: Electrons became fractions of themselves upon applying a current to a new device composed of rhombohedral pentalayer graphene and hexagonal boron nitride. This is important because this “fractional quantum Hall effect” has only been seen in a few systems, usually under very high magnetic fields. The Ju work showed that the phenomenon could occur in a fairly simple material without a magnetic field. As a result, it is called the “fractional quantum anomalous Hall effect” (anomalous indicates that no magnetic field is necessary).

New results

In the current work, the Ju team reports yet more unexpected phenomena from the general rhombohedral graphene/boron nitride system when it is cooled to 30 millikelvins (1 millikelvin is equivalent to -459.668 degrees Fahrenheit). In last year’s paper, Ju and colleagues reported six fractional states of electrons. In the current work, they report discovering two more of these fractional states.

They also found another unusual electronic phenomenon: the integer quantum anomalous Hall effect in a wide range of electron densities. The fractional quantum anomalous Hall effect was understood to emerge in an electron “liquid” phase, analogous to water. In contrast, the new state that the team has now observed can be interpreted as an electron “solid” phase — resembling the formation of electronic “ice” — that can also coexist with the fractional quantum anomalous Hall states when the system’s voltage is carefully tuned at ultra-low temperatures.

One way to think about the relation between the integer and fractional states is to imagine a map created by tuning electric voltages: By tuning the system with different voltages, you can create a “landscape” similar to a river (which represents the liquid-like fractional states) cutting through glaciers (which represent the solid-like integer effect), Ju explains.

Ju notes that his team observed all of these phenomena not only in pentalayer rhombohedral graphene, but also in rhombohedral graphene composed of four layers. This creates a family of materials, and indicates that other “relatives” may exist.

“This work shows how rich this material is in exhibiting exotic phenomena. We’ve just added more flavor to this already very interesting material,” says Zhengguang Lu, a co-first author of the paper. Lu, who conducted the work as a postdoc at MIT, is now on the faculty at Florida State University.

In addition to Ju and Lu, other principal authors of the Nature paper are Tonghang Han and Yuxuan Yao, both of MIT. Lu, Han, and Yao are co-first authors of the paper who contributed equally to the work. Other MIT authors are Jixiang Yang, Junseok Seo, Lihan Shi, and Shenyong Ye. Additional members of the team are Kenji Watanabe and Takashi Taniguchi of the National Institute for Materials Science in Japan.

This work was supported by a Sloan Fellowship, a Mathworks Fellowship, the U.S. Department of Energy, the Japan Society for the Promotion of Science KAKENHI, and the World Premier International Research Initiative of Japan. Device fabrication was performed at the Harvard Center for Nanoscale Systems and MIT.nano. 

What if the clothes you wear could care for your health?

MIT researchers have developed an autonomous programmable computer in the form of an elastic fiber, which could monitor health conditions and physical activity, alerting the wearer to potential health risks in real-time. Clothing containing the fiber computer was comfortable and machine washable, and the fibers were nearly imperceptible to the wearer, the researchers report.

Unlike on-body monitoring systems known as “wearables,” which are located at a single point like the chest, wrist, or finger, fabrics and apparel have an advantage of being in contact with large areas of the body close to vital organs. As such, they present a unique opportunity to measure and understand human physiology and health.

The fiber computer contains a series of microdevices, including sensors, a microcontroller, digital memory, bluetooth modules, optical communications, and a battery, making up all the necessary components of a computer in a single elastic fiber.

The researchers added four fiber computers to a top and a pair of leggings, with the fibers running along each limb. In their experiments, each independently programmable fiber computer operated a machine-learning model that was trained to autonomously recognize exercises performed by the wearer, resulting in an average accuracy of about 70 percent.

Surprisingly, once the researchers allowed the individual fiber computers to communicate among themselves, their collective accuracy increased to nearly 95 percent.

“Our bodies broadcast gigabytes of data through the skin every second in the form of heat, sound, biochemicals, electrical potentials, and light, all of which carry information about our activities, emotions, and health. Unfortunately, most — if not all — of it gets absorbed and then lost in the clothes we wear. Wouldn’t it be great if we could teach clothes to capture, analyze, store, and communicate this important information in the form of valuable health and activity insights?” says Yoel Fink, a professor of materials science and engineering at MIT, a principal investigator in the Research Laboratory of Electronics (RLE) and the Institute for Soldier Nanotechnologies (ISN), and senior author of a paper on the research, which appears today in Nature.

The use of the fiber computer to understand health conditions and help prevent injury will soon undergo a significant real-world test as well. U.S. Army and Navy service members will be conducting a monthlong winter research mission to the Arctic, covering 1,000 kilometers in average temperatures of -40 degrees Fahrenheit. Dozens of base layer merino mesh shirts with fiber computers will be providing real-time information on the health and activity of the individuals participating on this mission, called Musk Ox II.

“In the not-too-distant future, fiber computers will allow us to run apps and get valuable health care and safety services from simple everyday apparel. We are excited to see glimpses of this future in the upcoming Arctic mission through our partners in the U.S. Army, Navy, and DARPA. Helping to keep our service members safe in the harshest environments is a honor and privilege,” Fink says.

He is joined on the paper by co-lead authors Nikhil Gupta, an MIT materials science and engineering graduate student; Henry Cheung MEng ’23; and Syamantak Payra ’22, currently a graduate student at Stanford University; John Joannopoulos, the Francis Wright Professor of Physics at MIT and director of the Institute for Soldier Nanotechnologies; as well as others at MIT, Rhode Island School of Design, and Brown University.

Fiber focus

The fiber computer builds on more than a decade of work in the Fibers@MIT lab at the RLE and was supported primarily by ISN. In previous papers, the researchers demonstrated methods for incorporating semiconductor devices, optical diodes, memory units, elastic electrical contacts, and sensors into fibers that could be formed into fabrics and garments.

“But we hit a wall in terms of the complexity of the devices we could incorporate into the fiber because of how we were making it. We had to rethink the whole process. At the same time, we wanted to make it elastic and flexible so it would match the properties of traditional fabrics,” says Gupta.

One of the challenges that researchers surmounted is the geometric mismatch between a cylindrical fiber and a planar chip. Connecting wires to small, conductive areas, known as pads, on the outside of each planar microdevice proved to be difficult and prone to failure because complex microdevices have many pads, making it increasingly difficult to find room to attach each wire reliably.

In this new design, the researchers map the 2D pad alignment of each microdevice to a 3D layout using a flexible circuit board called an interposer, which they wrapped into a cylinder. They call this the “maki” design. Then, they attach four separate wires to the sides of the “maki” roll and connected all the components together.

“This advance was crucial for us in terms of being able to incorporate higher functionality computing elements, like the microcontroller and Bluetooth sensor, into the fiber,” says Gupta.

This versatile folding technique could be used with a variety of microelectronic devices, enabling them to incorporate additional functionality.

In addition, the researchers fabricated the new fiber computer using a type of thermoplastic elastomer that is several times more flexible than the thermoplastics they used previously. This material enabled them to form a machine-washable, elastic fiber that can stretch more than 60 percent without failure.

They fabricate the fiber computer using a thermal draw process that the Fibers@MIT group pioneered in the early 2000s. The process involves creating a macroscopic version of the fiber computer, called a preform, that contains each connected microdevice.

This preform is hung in a furnace, melted, and pulled down to form a fiber, which also contains embedded lithium-ion batteries so it can power itself.

“A former group member, Juliette Marion, figured out how to create elastic conductors, so even when you stretch the fiber, the conductors don’t break. We can maintain functionality while stretching it, which is crucial for processes like knitting, but also for clothes in general,” Gupta says.

Bring out the vote

Once the fiber computer is fabricated, the researchers use a braiding technique to cover the fiber with traditional yarns, such as polyester, merino wool, nylon, and even silk.

In addition to gathering data on the human body using sensors, each fiber computer incorporates LEDs and light sensors that enable multiple fibers in one garment to communicate, creating a textile network that can perform computation.

Each fiber computer also includes a Bluetooth communication system to send data wirelessly to a device like a smartphone, which can be read by a user.

The researchers leveraged these communication systems to create a textile network by sewing four fiber computers into a garment, one in each sleeve. Each fiber ran an independent neural network that was trained to identify exercises like squats, planks, arm circles, and lunges.

“What we found is that the ability of a fiber computer to identify human activity was only about 70 percent accurate when located on a single limb, the arms or legs. However, when we allowed the fibers sitting on all four limbs to ‘vote,’ they collectively reached nearly 95 percent accuracy, demonstrating the importance of residing on multiple body areas and forming a network between autonomous fiber computers that does not need wires and interconnects,” Fink says.

Moving forward, the researchers want to use the interposer technique to incorporate additional microdevices.

Arctic insights

In February, a multinational team equipped with computing fabrics will travel for 30 days and 1,000 kilometers in the Arctic. The fabrics will help keep the team safe, and set the stage for future physiological “digital twinning” models.

“As a leader with more than a decade of Arctic operational experience, one of my main concerns is how to keep my team safe from debilitating cold weather injuries — a primary threat to operators in the extreme cold,” says U.S. Army Major Mathew Hefner, the commander of Musk Ox II. “Conventional systems just don’t provide me with a complete picture. We will be wearing the base layer computing fabrics on us 24/7 to help us better understand the body’s response to extreme cold and ultimately predict and prevent injury.”

Karl Friedl, U.S. Army Research Institute of Environmental Medicine senior research scientist of performance physiology, noted that the MIT programmable computing fabric technology may become a “gamechanger for everyday lives.”

“Imagine near-term fiber computers in fabrics and apparel that sense and respond to the environment and to the physiological status of the individual, increasing comfort and performance, providing real-time health monitoring and providing protection against external threats. Soldiers will be the early adopters and beneficiaries of this new technology, integrated with AI systems using predictive physiological models and mission-relevant tools to enhance survivability in austere environments,” Friedl says.

“The convergence of classical fibers and fabrics with computation and machine learning has only begun. We are exploring this exciting future not only through research and field testing, but importantly in an MIT Department of Materials Science and Engineering course ‘Computing Fabrics,’ taught with Professor Anais Missakian from the Rhode Island School of Design,” adds Fink.

This research was supported, in part, by the U.S. Army Research Office Institute for Soldier Nanotechnology (ISN), the U.S. Defense Threat Reduction Agency, the U.S. National Science Foundation, the Fannie and John Hertz Foundation Fellowship, the Paul and Daisy Soros Foundation Fellowship for New Americans, the Stanford-Knight Hennessy Scholars Program, and the Astronaut Scholarship Foundation.

About 60 percent of all cancer patients in the United States receive radiation therapy as part of their treatment. However, this radiation can have severe side effects that often end up being too difficult for patients to tolerate.

Drawing inspiration from a tiny organism that can withstand huge amounts of radiation, researchers at MIT, Brigham and Women’s Hospital, and the University of Iowa have developed a new strategy that may protect patients from this kind of damage. Their approach makes use of a protein from tardigrades, often also called “water bears,” which are usually less than a millimeter in length.

When the researchers injected messenger RNA encoding this protein into mice, they found that it generated enough protein to protect cells’ DNA from radiation-induced damage. If developed for use in humans, this approach could benefit many cancer patients, the researchers say.

“Radiation can be very helpful for many tumors, but we also recognize that the side effects can be limiting. There’s an unmet need with respect to helping patients mitigate the risk of damaging adjacent tissue,” says Giovanni Traverso, an associate professor of mechanical engineering at MIT and a gastroenterologist at Brigham and Women’s Hospital.

Traverso and James Byrne, an assistant professor of radiation oncology at the University of Iowa, are the senior authors of the study, which appears today in Nature Biomedical Engineering. The paper’s lead authors are Ameya Kirtane, an instructor in medicine at Harvard Medical School and a visiting scientist at MIT’s Koch Institute for Integrative Cancer Research, and Jianling Bi, a research scientist at the University of Iowa.

Extreme survival

Radiation is often used to treat cancers of the head and neck, where it can damage the mouth or throat, making it very painful to eat or drink. It is also commonly used for gastrointestinal cancers, which can lead to rectal bleeding. Many patients end up delaying treatments or stopping them altogether.

“This affects a huge number of patients, and it can manifest as something as simple as mouth sores, which can limit a person’s ability to eat because it’s so painful, to requiring hospitalization because people are suffering so terribly from the pain, weight loss, or bleeding. It can be pretty dangerous, and it’s something that we really wanted to try and address,” Byrne says.

Currently, there are very few ways to prevent radiation damage in cancer patients. There are a handful of drugs that can be given to try to reduce the damage, and for prostate cancer patients, a hydrogel can be used to create a physical barrier between the prostate and the rectum during radiation treatment.

For several years, Traverso and Byrne have been working on developing new ways to prevent radiation damage. In the new study, they were inspired by the extraordinary survival ability of tardigrades. Found all over the world, usually in aquatic environments, these organisms are well known for their resilience to extreme conditions. Scientists have even sent them into space, where they were shown to survive extreme dehydration and cosmic radiation.

One key component of tardigrades’ defense systems is a unique damage suppressor protein called Dsup, which binds to DNA and helps protect it from radiation-induced damage. This protein plays a major role in tardigrades’ ability to survive radiation doses 2,000 to 3,000 times higher than what a human being can tolerate.

When brainstorming ideas for novel ways to protect cancer patients from radiation, the researchers wondered if they might be able to deliver messenger RNA encoding Dsup to patient tissues before radiation treatment. This mRNA would trigger cells to transiently express the protein, protecting DNA during the treatment. After a few hours, the mRNA and protein would disappear.

For this to work, the researchers needed a way to deliver mRNA that would generate large amounts of protein in the target tissues. They screened libraries of delivery particles containing both polymer and lipid components, which have been used separately to achieve efficient mRNA delivery. From these screens, they identified one polymer-lipid particle that was best-suited for delivery to the colon, and another that was optimized to deliver mRNA to mouth tissue.

“We thought that perhaps by combining these two systems — polymers and lipids — we may be able to get the best of both worlds and get highly potent RNA delivery. And that’s essentially what we saw,” Kirtane says. “One of the strengths of our approach is that we are using a messenger RNA, which just temporarily expresses the protein, so it’s considered far safer than something like DNA, which may be incorporated into the cells’ genome.”

Protection from radiation

After showing that these particles could successfully deliver mRNA to cells grown in the lab, the researchers tested whether this approach could effectively protect tissue from radiation in a mouse model.

They injected the particles into either the cheek or the rectum several hours before giving a dose of radiation similar to what cancer patients would receive. In these mice, the researchers saw a 50 percent reduction in the amount of double-stranded DNA breaks caused by radiation.

“This study shows great promise and is a really novel idea leveraging natural mechanisms of protection again DNA damage for the purpose of protecting healthy cells during radiation treatments for cancer,” says Ben Ho Park, director of the Vanderbilt-Ingram Cancer Center at Vanderbilt University Medical Center, who was not involved in the study.

The researchers also showed that the protective effect of the Dsup protein did not spread beyond the injection site, which is important because they don’t want to protect the tumor itself from the effects of radiation. To make this treatment more feasible for potential use in humans, the researchers now plan to work on developing a version of the Dsup protein that would not provoke an immune response, as the original tardigrade protein likely would.

If developed for use in humans, this protein could also potentially be used to protect against DNA damage caused by chemotherapy drugs, the researchers say. Another possible application would be to help prevent radiation damage in astronauts in space.

Other authors of the paper include Netra Rajesh, Chaoyang Tang, Miguel Jimenez, Emily Witt, Megan McGovern, Arielle Cafi, Samual Hatfield, Lauren Rosenstock, Sarah Becker, Nicole Machado, Veena Venkatachalam, Dylan Freitas, Xisha Huang, Alvin Chan, Aaron Lopes, Hyunjoon Kim, Nayoon Kim, Joy Collins, Michelle Howard, Srija Manchkanti, and Theodore Hong.

The research was funded by the Prostate Cancer Foundation Young Investigator Award, the U.S. Department of Defense Prostate Cancer Program Early Investigator Award, a Hope Funds for Cancer Research Fellowship, the American Cancer Society, the National Cancer Institute, MIT’s Department of Mechanical Engineering, and the U.S. Advanced Research Projects Agency for Health. 

Maybe it’s a life hack or a liability, or a little of both. A surprising result in a new MIT study may suggest that people and animals alike share an inherent propensity to keep updating their approach to a task even when they have already learned how they should approach it, and even if the deviations sometimes lead to unnecessary error.

The behavior of “exploring” when one could just be “exploiting” could make sense for at least two reasons, says Mriganka Sur, senior author of the study published Feb. 18 in Current Biology. Just because a task’s rules seem set one moment doesn’t mean they’ll stay that way in this uncertain world, so altering behavior from the optimal condition every so often could help reveal needed adjustments. Moreover, trying new things when you already know what you like is a way of finding out whether there might be something even better out there than the good thing you’ve got going on right now.

“If the goal is to maximize reward, you should never deviate once you have found the perfect solution, yet you keep exploring,” says Sur, the Paul and Lilah Newton Professor in The Picower Institute for Learning and Memory and the Department of Brain and Cognitive Sciences at MIT. “Why? It’s like food. We all like certain foods, but we still keep trying different foods because you never know, there might be something you could discover.”

Predicting timing

Former research technician Tudor Dragoi, now a graduate student at Boston University, led the study in which he and fellow members of the Sur Lab explored how humans and marmosets, a small primate, make predictions about event timing.

Three humans and two marmosets were given a simple task. They’d see an image on a screen for some amount of time — the amount of time varied from one trial to the next within a limited range — and they simply had to hit a button (marmosets poked a tablet while humans clicked a mouse) when the image disappeared. Success was defined as reacting as quickly as possible to the image’s disappearance without hitting the button too soon. Marmosets received a juice reward on successful trials.

Though marmosets needed more training time than humans, the subjects all settled into the same reasonable pattern of behavior regarding the task. The longer the image stayed on the screen, the faster their reaction time to its disappearance. This behavior follows the “hazard model” of prediction in which, if the image can only last for so long, the longer it’s still there, the more likely it must be to disappear very soon. The subjects learned this and overall, with more experience, their reaction times became faster.

But as the experiment continued, Sur and Dragoi’s team noticed something surprising was also going on. Mathematical modeling of the reaction time data revealed that both the humans and marmosets were letting the results of the immediate previous trial influence what they did on the next trial, even though they had already learned what to do. If the image was only on the screen briefly in one trial, on the next round subjects would decrease reaction time a bit (presumably expecting a shorter image duration again) whereas if the image lingered, they’d increase reaction time (presumably because they figured they’d have a longer wait).

Those results add to ones from a similar study Sur’s lab published in 2023, in which they found that even after mice learned the rules of a different cognitive task, they’d arbitrarily deviate from the winning strategy every so often. In that study, like this one, learning the successful strategy didn’t prevent subjects from continuing to test alternatives, even if it meant sacrificing reward.

“The persistence of behavioral changes even after task learning may reflect exploration as a strategy for seeking and setting on an optimal internal model of the environment,” the scientists wrote in the new study.

Relevance for autism

The similarity of the human and marmoset behaviors is an important finding as well, Sur says. That’s because differences in making predictions about one’s environment is posited to be a salient characteristic of autism spectrum disorders. Because marmosets are small, are inherently social, and are more cognitively complex than mice, work has begun in some labs to establish marmoset autism models, but a key component was establishing that they model autism-related behaviors well. By demonstrating that marmosets model neurotypical human behavior regarding predictions, the study therefore adds weight to the emerging idea that marmosets can indeed provide informative models for autism studies.

In addition to Dragoi and Sur, other authors of the paper are Hiroki Sugihara, Nhat Le, Elie Adam, Jitendra Sharma, Guoping Feng, and Robert Desimone.

The Simons Foundation Autism Research Initiative supported the research through the Simons Center for the Social Brain at MIT.

Rain can freefall at speeds of up to 25 miles per hour. If the droplets land in a puddle or pond, they can form a crown-like splash that, with enough force, can dislodge any surface particles and launch them into the air.

Now MIT scientists have taken high-speed videos of droplets splashing into a deep pool, to track how the fluid evolves, above and below the water line, frame by millisecond frame. Their work could help to predict how spashing droplets, such as from rainstorms and irrigation systems, may impact watery surfaces and aerosolize surface particles, such as pollen on puddles or pesticides in agricultural runoff.

The team carried out experiments in which they dispensed water droplets of various sizes and from various heights into a pool of water. Using high-speed imaging, they measured how the liquid pool deformed as the impacting droplet hit the pool’s surface.

Across all their experiments, they observed a common splash evolution: As a droplet hit the pool, it pushed down below the surface to form a “crater,” or cavity. At nearly the same time, a wall of liquid rose above the surface, forming a crown. Interestingly, the team observed that small, secondary droplets were ejected from the crown before the crown reached its maximum height. This entire evolution happens in a fraction of a second.

Scientists have caught snapshots of droplet splashes in the past, such as the famous “Milk Drop Coronet” — a photo of a drop of milk in mid-splash, taken by the late MIT professor Harold “Doc” Edgerton, who invented a photographic technique to capture quickly moving objects.

The new work represents the first time scientists have used such high-speed images to model the entire splash dynamics of a droplet in a deep pool, combining what happens both above and below the surface. The team has used the imaging to gather new data central to build a mathematical model that predicts how a droplet’s shape will morph and merge as it hits a pool’s surface. They plan to use the model as a baseline to explore to what extent a splashing droplet might drag up and launch particles from the water pool.

“Impacts of drops on liquid layers are ubiquitous,” says study author Lydia Bourouiba, a professor in the MIT departments of Civil and Environmental Engineering and Mechanical Engineering, and a core member of the Institute for Medical Engineering and Science (IMES). “Such impacts can produce myriads of secondary droplets that could act as carriers for pathogens, particles, or microbes that are on the surface of impacted pools or contaminated water bodies. This work is key in enabling prediction of droplet size distributions, and potentially also what such drops can carry with them.”

Bourouiba and her mentees have published their results in the Journal of Fluid Mechanics. MIT co-authors include former graduate student Raj Dandekar PhD ’22, postdoc (Eric) Naijian Shen, and student mentee Boris Naar.

Above and below

At MIT, Bourouiba heads up the Fluid Dynamics of Disease Transmission Laboratory, part of the Fluids and Health Network, where she and her team explore the fundamental physics of fluids and droplets in a range of environmental, energy, and health contexts, including disease transmission. For their new study, the team looked to better understand how droplets impact a deep pool — a seemingly simple phenomenon that nevertheless has been tricky to precisely capture and characterize.

Bourouiba notes that there have been recent breakthroughs in modeling the evolution of a splashing droplet below a pool’s surface. As a droplet hits a pool of water, it breaks through the surface and drags air down through the pool to create a short-lived crater. Until now, scientists have focused on the evolution of this underwater cavity, mainly for applications in energy harvesting. What happens above the water, and how a droplet’s crown-like shape evolves with the cavity below, remained less understood.

“The descriptions and understanding of what happens below the surface, and above, have remained very much divorced,” says Bourouiba, who believes such an understanding can help to predict how droplets launch and spread chemicals, particles, and microbes into the air.

Splash in 3D

To study the coupled dynamics between a droplet’s cavity and crown, the team set up an experiment to dispense water droplets into a deep pool. For the purposes of their study, the researchers considered a deep pool to be a body of water that is deep enough that a splashing droplet would remain far away from the pool’s bottom. In these terms, they found that a pool with a depth of at least 20 centimeters was sufficient for their experiments.

They varied each droplet’s size, with an average diameter of about 5 millimeters. They also dispensed droplets from various heights, causing the droplets to hit the pool’s surface at different speeds, which on average was about 5 meters per second. The overall dynamics, Bourouiba says, should be similar to what occurs on the surface of a puddle or pond during an average rainstorm.

“This is capturing the speed at which raindrops fall,” she says. “These wouldn’t be very small, misty drops. This would be rainstorm drops for which one needs an umbrella.”

Using high-speed imaging techniques inspired by Edgerton’s pioneering photography, the team captured videos of pool-splashing droplets, at rates of up to 12,500 frames per second. They then applied in-house imaging processing methods to extract key measurements from the image sequences, such as the changing width and depth of the underwater cavity, and the evolving diameter and height of the rising crown. The researchers also captured especially tricky measurements, of the crown’s wall thickness profile and inner flow — the cylinder that rises out of the pool, just before it forms a rim and points that are characteristic of a crown.

“This cylinder-like wall of rising liquid, and how it evolves in time and space, is at the heart of everything,” Bourouiba says. “It’s what connects the fluid from the pool to what will go into the rim and then be ejected into the air through smaller, secondary droplets.”

The researchers worked the image data into a set of “evolution equations,” or a mathematical model that relates the various properties of an impacting droplet, such as the width of its cavity and the thickness and speed profiles of its crown wall, and how these properties change over time, given a droplet’s starting size and impact speed.

“We now have a closed-form mathematical expression that people can use to see how all these quantities of a splashing droplet change over space and time,” says co-author Shen, who plans, with Bourouiba, to apply the new model to the behavior of secondary droplets and understanding how a splash end-up dispersing particles such as pathogens and pesticides. “This opens up the possibility to study all these problems of splash in 3D, with self-contained closed-formed equations, which was not possible before.”

This research was supported, in part, by the Department of Agriculture-National Institute of Food and Agriculture Specialty Crop Research Initiative; the Richard and Susan Smith Family Foundation; the National Science Foundation; the Centers for Disease Control and Prevention-National Institute for Occupational Safety and Health; Inditex; and the National Institute of Allergy and Infectious Diseases of the National Institutes of Health.

As part of a multi-pronged approach toward curbing the effects of greenhouse gas emissions, scientists seek to better understand the impact of rising carbon dioxide (CO₂) levels on terrestrial ecosystems, particularly tropical forests. To that end, climate scientist César Terrer, the Class of 1958 Career Development Assistant Professor of Civil and Environmental Engineering (CEE) at MIT, and colleague Josh Fisher of Chapman University are bringing their scientific minds to bear on a unique setting — an active volcano in Costa Rica — as a way to study carbon dioxide emissions and their influence. 

Elevated CO₂ levels can lead to a phenomenon known as the CO₂ fertilization effect, where plants grow more and absorb greater amounts of carbon, providing a cooling effect. While this effect has the potential to be a natural climate change mitigator, the extent of how much carbon plants can continue to absorb remains uncertain. There are growing concerns from scientists that plants may eventually reach a saturation point, losing their ability to offset increasing atmospheric CO₂. Understanding these dynamics is crucial for accurate climate predictions and developing strategies to manage carbon sequestration. Here, Terrer discusses his innovative approach, his motivations for joining the project, and the importance of advancing this research.

Q: Why did you get involved in this line of research, and what makes it unique?

A: Josh Fisher, a climate scientist and long-time collaborator, had the brilliant idea to take advantage of naturally high CO₂ levels near active volcanoes to study the fertilization effect in real-world conditions. Conducting such research in dense tropical forests like the Amazon — where the largest uncertainties about CO₂ fertilization exist — is challenging. It would require large-scale CO₂ tanks and extensive infrastructure to evenly distribute the gas throughout the towering trees and intricate canopy layers — a task that is not only logistically complex, but also highly costly. Our approach allows us to circumvent those obstacles and gather critical data in a way that hasn't been done before.

Josh was looking for an expert in the field of carbon ecology to co-lead and advance this research with him. My expertise of understanding the dynamics that regulate carbon storage in terrestrial ecosystems within the context of climate change made for a natural fit to co-lead and advance this research with him. This field has been central to my research, and was the focus of my PhD thesis.

Our experiments inside the Rincon de la Vieja National Park are particularly exciting because CO₂ concentrations in the areas near the volcano are four times higher than the global average. This gives us a rare opportunity to observe how elevated CO₂ affects plant biomass in a natural setting — something that has never been attempted at this scale.

Q: How are you measuring CO₂ concentrations at the volcano?

A: We have installed a network of 50 sensors in the forest canopy surrounding the volcano. These sensors continuously monitor CO₂ levels, allowing us to compare areas with naturally high CO₂ emissions from the volcano to control areas with typical atmospheric CO₂ concentrations. The sensors are Bluetooth-enabled, requiring us to be in close proximity to retrieve the data. They will remain in place for a full year, capturing a continuous dataset on CO₂ fluctuations. Our next data collection trip is scheduled for March, with another planned a year after the initial deployment.

Q: What are the long-term goals of this research?

A: Our primary objective is to determine whether the CO₂ fertilization effect can be sustained, or if plants will eventually reach a saturation point, limiting their ability to absorb additional carbon. Understanding this threshold is crucial for improving climate models and carbon mitigation strategies.

To expand the scope of our measurements, we are exploring the use of airborne technologies — such as drones or airplane-mounted sensors — to assess carbon storage across larger areas. This would provide a more comprehensive view of carbon sequestration potential in tropical ecosystems. Ultimately, this research could offer critical insights into the future role of forests in mitigating climate change, helping scientists and policymakers develop more accurate carbon budgets and climate projections. If successful, our approach could pave the way for similar studies in other ecosystems, deepening our understanding of how nature responds to rising CO₂ levels.

All biological function is dependent on how different proteins interact with each other. Protein-protein interactions facilitate everything from transcribing DNA and controlling cell division to higher-level functions in complex organisms.

Much remains unclear, however, about how these functions are orchestrated on the molecular level, and how proteins interact with each other — either with other proteins or with copies of themselves.

Recent findings have revealed that small protein fragments have a lot of functional potential. Even though they are incomplete pieces, short stretches of amino acids can still bind to interfaces of a target protein, recapitulating native interactions. Through this process, they can alter that protein’s function or disrupt its interactions with other proteins.

Protein fragments could therefore empower both basic research on protein interactions and cellular processes, and could potentially have therapeutic applications.

Recently published in Proceedings of the National Academy of Sciences, a new method developed in the Department of Biology builds on existing artificial intelligence models to computationally predict protein fragments that can bind to and inhibit full-length proteins in E. coli. Theoretically, this tool could lead to genetically encodable inhibitors against any protein.

The work was done in the lab of associate professor of biology and Howard Hughes Medical Institute investigator Gene-Wei Li in collaboration with the lab of Jay A. Stein (1968) Professor of Biology, professor of biological engineering, and department head Amy Keating.

Leveraging machine learning

The program, called FragFold, leverages AlphaFold, an AI model that has led to phenomenal advancements in biology in recent years due to its ability to predict protein folding and protein interactions.

The goal of the project was to predict fragment inhibitors, which is a novel application of AlphaFold. The researchers on this project confirmed experimentally that more than half of FragFold’s predictions for binding or inhibition were accurate, even when researchers had no previous structural data on the mechanisms of those interactions.

“Our results suggest that this is a generalizable approach to find binding modes that are likely to inhibit protein function, including for novel protein targets, and you can use these predictions as a starting point for further experiments,” says co-first and corresponding author Andrew Savinov, a postdoc in the Li Lab. “We can really apply this to proteins without known functions, without known interactions, without even known structures, and we can put some credence in these models we’re developing.”

One example is FtsZ, a protein that is key for cell division. It is well-studied but contains a region that is intrinsically disordered and, therefore, especially challenging to study. Disordered proteins are dynamic, and their functional interactions are very likely fleeting — occurring so briefly that current structural biology tools can’t capture a single structure or interaction.

The researchers leveraged FragFold to explore the activity of fragments of FtsZ, including fragments of the intrinsically disordered region, to identify several new binding interactions with various proteins. This leap in understanding confirms and expands upon previous experiments measuring FtsZ’s biological activity.

This progress is significant in part because it was made without solving the disordered region’s structure, and because it exhibits the potential power of FragFold.

“This is one example of how AlphaFold is fundamentally changing how we can study molecular and cell biology,” Keating says. “Creative applications of AI methods, such as our work on FragFold, open up unexpected capabilities and new research directions.”

Inhibition, and beyond

The researchers accomplished these predictions by computationally fragmenting each protein and then modeling how those fragments would bind to interaction partners they thought were relevant.

They compared the maps of predicted binding across the entire sequence to the effects of those same fragments in living cells, determined using high-throughput experimental measurements in which millions of cells each produce one type of protein fragment.

AlphaFold uses co-evolutionary information to predict folding, and typically evaluates the evolutionary history of proteins using something called multiple sequence alignments for every single prediction run. The MSAs are critical, but are a bottleneck for large-scale predictions — they can take a prohibitive amount of time and computational power.

For FragFold, the researchers instead pre-calculated the MSA for a full-length protein once, and used that result to guide the predictions for each fragment of that full-length protein.

Savinov, together with Keating Lab alumnus Sebastian Swanson PhD ’23, predicted inhibitory fragments of a diverse set of proteins in addition to FtsZ. Among the interactions they explored was a complex between lipopolysaccharide transport proteins LptF and LptG. A protein fragment of LptG inhibited this interaction, presumably disrupting the delivery of lipopolysaccharide, which is a crucial component of the E. coli outer cell membrane essential for cellular fitness.

“The big surprise was that we can predict binding with such high accuracy and, in fact, often predict binding that corresponds to inhibition,” Savinov says. “For every protein we’ve looked at, we’ve been able to find inhibitors.”

The researchers initially focused on protein fragments as inhibitors because whether a fragment could block an essential function in cells is a relatively simple outcome to measure systematically. Looking forward, Savinov is also interested in exploring fragment function outside inhibition, such as fragments that can stabilize the protein they bind to, enhance or alter its function, or trigger protein degradation.

Design, in principle

This research is a starting point for developing a systemic understanding of cellular design principles, and what elements deep-learning models may be drawing on to make accurate predictions.

“There’s a broader, further-reaching goal that we’re building towards,” Savinov says. “Now that we can predict them, can we use the data we have from predictions and experiments to pull out the salient features to figure out what AlphaFold has actually learned about what makes a good inhibitor?”

Savinov and collaborators also delved further into how protein fragments bind, exploring other protein interactions and mutating specific residues to see how those interactions change how the fragment interacts with its target.

Experimentally examining the behavior of thousands of mutated fragments within cells, an approach known as deep mutational scanning, revealed key amino acids that are responsible for inhibition. In some cases, the mutated fragments were even more potent inhibitors than their natural, full-length sequences.

“Unlike previous methods, we are not limited to identifying fragments in experimental structural data,” says Swanson. “The core strength of this work is the interplay between high-throughput experimental inhibition data and the predicted structural models: the experimental data guides us towards the fragments that are particularly interesting, while the structural models predicted by FragFold provide a specific, testable hypothesis for how the fragments function on a molecular level.”

Savinov is excited about the future of this approach and its myriad applications.

“By creating compact, genetically encodable binders, FragFold opens a wide range of possibilities to manipulate protein function,” Li agrees. “We can imagine delivering functionalized fragments that can modify native proteins, change their subcellular localization, and even reprogram them to create new tools for studying cell biology and treating diseases.” 

There’s a lot of untapped potential in our homes and vehicles that could be harnessed to reinforce local power grids and make them more resilient to unforeseen outages, a new study shows.

In response to a cyber attack or natural disaster, a backup network of decentralized devices — such as residential solar panels, batteries, electric vehicles, heat pumps, and water heaters — could restore electricity or relieve stress on the grid, MIT engineers say.

Such devices are “grid-edge” resources found close to the consumer rather than near central power plants, substations, or transmission lines. Grid-edge devices can independently generate, store, or tune their consumption of power. In their study, the research team shows how such devices could one day be called upon to either pump power into the grid, or rebalance it by dialing down or delaying their power use.

In a paper appearing this week in the Proceedings of the National Academy of Sciences, the engineers present a blueprint for how grid-edge devices could reinforce the power grid through a “local electricity market.” Owners of grid-edge devices could subscribe to a regional market and essentially loan out their device to be part of a microgrid or a local network of on-call energy resources.

In the event that the main power grid is compromised, an algorithm developed by the researchers would kick in for each local electricity market, to quickly determine which devices in the network are trustworthy. The algorithm would then identify the combination of trustworthy devices that would most effectively mitigate the power failure, by either pumping power into the grid or reducing the power they draw from it, by an amount that the algorithm would calculate and communicate to the relevant subscribers. The subscribers could then be compensated through the market, depending on their participation.

The team illustrated this new framework through a number of grid attack scenarios, in which they considered failures at different levels of a power grid, from various sources such as a cyber attack or a natural disaster. Applying their algorithm, they showed that various networks of grid-edge devices were able to dissolve the various attacks.

The results demonstrate that grid-edge devices such as rooftop solar panels, EV chargers, batteries, and smart thermostats (for HVAC devices or heat pumps) could be tapped to stabilize the power grid in the event of an attack.

“All these small devices can do their little bit in terms of adjusting their consumption,” says study co-author Anu Annaswamy, a research scientist in MIT’s Department of Mechanical Engineering. “If we can harness our smart dishwashers, rooftop panels, and EVs, and put our combined shoulders to the wheel, we can really have a resilient grid.”

The study’s MIT co-authors include lead author Vineet Nair and John Williams, along with collaborators from multiple institutions including the Indian Institute of Technology, the National Renewable Energy Laboratory, and elsewhere.

Power boost

The team’s study is an extension of their broader work in adaptive control theory and designing systems to automatically adapt to changing conditions. Annaswamy, who leads the Active-Adaptive Control Laboratory at MIT, explores ways to boost the reliability of renewable energy sources such as solar power.

“These renewables come with a strong temporal signature, in that we know for sure the sun will set every day, so the solar power will go away,” Annaswamy says. “How do you make up for the shortfall?”

The researchers found the answer could lie in the many grid-edge devices that consumers are increasingly installing in their own homes.

“There are lots of distributed energy resources that are coming up now, closer to the customer rather than near large power plants, and it’s mainly because of individual efforts to decarbonize,” Nair says. “So you have all this capability at the grid edge. Surely we should be able to put them to good use.”

While considering ways to deal with drops in energy from the normal operation of renewable sources, the team also began to look into other causes of power dips, such as from cyber attacks. They wondered, in these malicious instances, whether and how the same grid-edge devices could step in to stabilize the grid following an unforeseen, targeted attack.

Attack mode

In their new work, Annaswamy, Nair, and their colleagues developed a framework for incorporating grid-edge devices, and in particular, internet-of-things (IoT) devices, in a way that would support the larger grid in the event of an attack or disruption. IoT devices are physical objects that contain sensors and software that connect to the internet.

For their new framework, named EUREICA (Efficient, Ultra-REsilient, IoT-Coordinated Assets), the researchers start with the assumption that one day, most grid-edge devices will also be IoT devices, enabling rooftop panels, EV chargers, and smart thermostats to wirelessly connect to a larger network of similarly independent and distributed devices. 

The team envisions that for a given region, such as a community of 1,000 homes, there exists a certain number of IoT devices that could potentially be enlisted in the region’s local network, or microgrid. Such a network would be managed by an operator, who would be able to communicate with operators of other nearby microgrids.

If the main power grid is compromised or attacked, operators would run the researchers’ decision-making algorithm to determine trustworthy devices within the network that can pitch in to help mitigate the attack.

The team tested the algorithm on a number of scenarios, such as a cyber attack in which all smart thermostats made by a certain manufacturer are hacked to raise their setpoints simultaneously to a degree that dramatically alters a region’s energy load and destabilizes the grid. The researchers also considered attacks and weather events that would shut off the transmission of energy at various levels and nodes throughout a power grid.

“In our attacks we consider between 5 and 40 percent of the power being lost. We assume some nodes are attacked, and some are still available and have some IoT resources, whether a battery with energy available or an EV or HVAC device that’s controllable,” Nair explains. “So, our algorithm decides which of those houses can step in to either provide extra power generation to inject into the grid or reduce their demand to meet the shortfall.”

In every scenario that they tested, the team found that the algorithm was able to successfully restabilize the grid and mitigate the attack or power failure. They acknowledge that to put in place such a network of grid-edge devices will require buy-in from customers, policymakers, and local officials, as well as innovations such as advanced power inverters that enable EVs to inject power back into the grid.

“This is just the first of many steps that have to happen in quick succession for this idea of local electricity markets to be implemented and expanded upon,” Annaswamy says. “But we believe it’s a good start.”

This work was supported, in part, by the U.S. Department of Energy and the MIT Energy Initiative.

RNA splicing is a cellular process that is critical for gene expression. After genes are copied from DNA into messenger RNA, portions of the RNA that don’t code for proteins, called introns, are cut out and the coding portions are spliced back together.

This process is controlled by a large protein-RNA complex called the spliceosome. MIT biologists have now discovered a new layer of regulation that helps to determine which sites on the messenger RNA molecule the spliceosome will target.

The research team discovered that this type of regulation, which appears to influence the expression of about half of all human genes, is found throughout the animal kingdom, as well as in plants. The findings suggest that the control of RNA splicing, a process that is fundamental to gene expression, is more complex than previously known.

“Splicing in more complex organisms, like humans, is more complicated than it is in some model organisms like yeast, even though it’s a very conserved molecular process. There are bells and whistles on the human spliceosome that allow it to process specific introns more efficiently. One of the advantages of a system like this may be that it allows more complex types of gene regulation,” says Connor Kenny, an MIT graduate student and the lead author of the study.

Christopher Burge, the Uncas and Helen Whitaker Professor of Biology at MIT, is the senior author of the study, which appears today in Nature Communications.

Building proteins

RNA splicing, a process discovered in the late 1970s, allows cells to precisely control the content of the mRNA transcripts that carry the instructions for building proteins.

Each mRNA transcript contains coding regions, known as exons, and noncoding regions, known as introns. They also include sites that act as signals for where splicing should occur, allowing the cell to assemble the correct sequence for a desired protein. This process enables a single gene to produce multiple proteins; over evolutionary timescales, splicing can also change the size and content of genes and proteins, when different exons become included or excluded.

The spliceosome, which forms on introns, is composed of proteins and noncoding RNAs called small nuclear RNAs (snRNAs). In the first step of spliceosome assembly, an snRNA molecule known as U1 snRNA binds to the 5’ splice site at the beginning of the intron. Until now, it had been thought that the binding strength between the 5’ splice site and the U1 snRNA was the most important determinant of whether an intron would be spliced out of the mRNA transcript.

In the new study, the MIT team discovered that a family of proteins called LUC7 also helps to determine whether splicing will occur, but only for a subset of introns — in human cells, up to 50 percent.

Before this study, it was known that LUC7 proteins associate with U1 snRNA, but the exact function wasn’t clear. There are three different LUC7 proteins in human cells, and Kenny’s experiments revealed that two of these proteins interact specifically with one type of 5’ splice site, which the researchers called “right-handed.” A third human LUC7 protein interacts with a different type, which the researchers call “left-handed.”

The researchers found that about half of human introns contain a right- or left-handed site, while the other half do not appear to be controlled by interaction with LUC7 proteins. This type of control appears to add another layer of regulation that helps remove specific introns more efficiently, the researchers say.

“The paper shows that these two different 5’ splice site subclasses exist and can be regulated independently of one another,” Kenny says. “Some of these core splicing processes are actually more complex than we previously appreciated, which warrants more careful examination of what we believe to be true about these highly conserved molecular processes.”

“Complex splicing machinery”

Previous work has shown that mutation or deletion of one of the LUC7 proteins that bind to right-handed splice sites is linked to blood cancers, including about 10 percent of acute myeloid leukemias (AMLs). In this study, the researchers found that AMLs that lost a copy of the LUC7L2 gene have inefficient splicing of right-handed splice sites. These cancers also developed the same type of altered metabolism seen in earlier work.

“Understanding how the loss of this LUC7 protein in some AMLs alters splicing could help in the design of therapies that exploit these splicing differences to treat AML,” Burge says. “There are also small molecule drugs for other diseases such as spinal muscular atrophy that stabilize the interaction between U1 snRNA and specific 5’ splice sites. So the knowledge that particular LUC7 proteins influence these interactions at specific splice sites could aid in improving the specificity of this class of small molecules.”

Working with a lab led by Sascha Laubinger, a professor at Martin Luther University Halle-Wittenberg, the researchers found that introns in plants also have right- and left-handed 5’ splice sites that are regulated by Luc7 proteins.

The researchers’ analysis suggests that this type of splicing arose in a common ancestor of plants, animals, and fungi, but it was lost from fungi soon after they diverged from plants and animals.

“A lot what we know about how splicing works and what are the core components actually comes from relatively old yeast genetics work,” Kenny says. “What we see is that humans and plants tend to have more complex splicing machinery, with additional components that can regulate different introns independently.”

The researchers now plan to further analyze the structures formed by the interactions of Luc7 proteins with mRNA and the rest of the spliceosome, which could help them figure out in more detail how different forms of Luc7 bind to different 5’ splice sites.

The research was funded by the U.S. National Institutes of Health and the German Research Foundation.

The use of terahertz waves, which have shorter wavelengths and higher frequencies than radio waves, could enable faster data transmission, more precise medical imaging, and higher-resolution radar.

But effectively generating terahertz waves using a semiconductor chip, which is essential for incorporation into electronic devices, is notoriously difficult.

Many current techniques can’t generate waves with enough radiating power for useful applications unless they utilize bulky and expensive silicon lenses. Higher radiating power allows terahertz signals to travel farther. Such lenses, which are often larger than the chip itself, make it hard to integrate the terahertz source into an electronic device.

To overcome these limitations, MIT researchers developed a terahertz amplifier-multiplier system that achieves higher radiating power than existing devices without the need for silicon lenses.

By affixing a thin, patterned sheet of material to the back of the chip and utilizing higher-power Intel transistors, the researchers produced a more efficient, yet scalable, chip-based terahertz wave generator.

This compact chip could be used to make terahertz arrays for applications like improved security scanners for detecting hidden objects or environmental monitors for pinpointing airborne pollutants.

“To take full advantage of a terahertz wave source, we need it to be scalable. A terahertz array might have hundreds of chips, and there is no place to put silicon lenses because the chips are combined with such high density. We need a different package, and here we’ve demonstrated a promising approach that can be used for scalable, low-cost terahertz arrays,” says Jinchen Wang, a graduate student in the Department of Electrical Engineering and Computer Science (EECS) and lead author of a paper on the terahertz radiator.

He is joined on the paper by EECS graduate students Daniel Sheen and Xibi Chen; Steven F. Nagle, managing director of the T.J. Rodgers RLE Laboratory; and senior author Ruonan Han, an associate professor in EECS, who leads the Terahertz Integrated Electronics Group. The research will be presented at the IEEE International Solid-States Circuits Conference.

Making waves

Terahertz waves sit on the electromagnetic spectrum between radio waves and infrared light. Their higher frequencies enable them to carry more information per second than radio waves, while they can safely penetrate a wider range of materials than infrared light.

One way to generate terahertz waves is with a CMOS chip-based amplifier-multiplier chain that increases the frequency of radio waves until they reach the terahertz range. To achieve the best performance, waves go through the silicon chip and are eventually emitted out the back into the open air.

But a property known as the dielectric constant gets in the way of a smooth transmission.

The dielectric constant influences how electromagnetic waves interact with a material. It affects the amount of radiation that is absorbed, reflected, or transmitted. Because the dielectric constant of silicon is much higher than that of air, most terahertz waves are reflected at the silicon-air boundary rather than being cleanly transmitted out the back.

Since most signal strength is lost at this boundary, current approaches often use silicon lenses to boost the power of the remaining signal. 

The MIT researchers approached this problem differently.

They drew on an electromechanical theory known as matching. With matching, they seek to equal out the dielectric constants of silicon and air, which will minimize the amount of signal that is reflected at the boundary.

They accomplish this by sticking a thin sheet of material which has a dielectric constant between silicon and air to the back of the chip. With this matching sheet in place, most waves will be transmitted out the back rather than being reflected.

A scalable approach

They chose a low-cost, commercially available substrate material with a dielectric constant very close to what they needed for matching. To improve performance, they used a laser cutter to punch tiny holes into the sheet until its dielectric constant was exactly right.

“Since the dielectric constant of air is 1, if you just cut some subwavelength holes in the sheet, it is equivalent to injecting some air, which lowers the overall dielectric constant of the matching sheet,” Wang explains.

In addition, they designed their chip with special transistors developed by Intel that have a higher maximum frequency and breakdown voltage than traditional CMOS transistors.

“These two things taken together, the more powerful transistors and the dielectric sheet, plus a few other small innovations, enabled us to outperform several other devices,” he says.

Their chip generated terahertz signals with a peak radiation power of 11.1 decibel-milliwatts, the best among state-of-the-art techniques. Moreover, since the low-cost chip can be fabricated at scale, it could be integrated into real-world electronic devices more readily.

One of the biggest challenges of developing a scalable chip was determining how to manage the power and temperature when generating terahertz waves.

“Because the frequency and the power are so high, many of the standard ways to design a CMOS chip are not applicable here,” Wang says.

The researchers also needed to devise a technique for installing the matching sheet that could be scaled up in a manufacturing facility.

Moving forward, they want to demonstrate this scalability by fabricating a phased array of CMOS terahertz sources, enabling them to steer and focus a powerful terahertz beam with a low-cost, compact device.

This research is supported, in part, by NASA’s Jet Propulsion Laboratory and Strategic University Research Partnerships Program, as well as the MIT Center for Integrated Circuits and Systems. The chip was fabricated through the Intel University Shuttle Program.

In the race to reduce climate-warming carbon emissions, the buildings sector is falling behind. While carbon dioxide (CO₂) emissions in the U.S. electric power sector dropped by 34 percent between 2005 and 2021, emissions in the building sector declined by only 18 percent in that same time period. Moreover, in extremely cold locations, burning natural gas to heat houses can make up a substantial share of the emissions portfolio. Therefore, steps to electrify buildings in general, and residential heating in particular, are essential for decarbonizing the U.S. energy system.

But that change will increase demand for electricity and decrease demand for natural gas. What will be the net impact of those two changes on carbon emissions and on the cost of decarbonizing? And how will the electric power and natural gas sectors handle the new challenges involved in their long-term planning for future operations and infrastructure investments?

A new study by MIT researchers with support from the MIT Energy Initiative (MITEI) Future Energy Systems Center unravels the impacts of various levels of electrification of residential space heating on the joint power and natural gas systems. A specially devised modeling framework enabled them to estimate not only the added costs and emissions for the power sector to meet the new demand, but also any changes in costs and emissions that result for the natural gas sector.

The analyses brought some surprising outcomes. For example, they show that — under certain conditions — switching 80 percent of homes to heating by electricity could cut carbon emissions and at the same time significantly reduce costs over the combined natural gas and electric power sectors relative to the case in which there is only modest switching. That outcome depends on two changes: Consumers must install high-efficiency heat pumps plus take steps to prevent heat losses from their homes, and planners in the power and the natural gas sectors must work together as they make long-term infrastructure and operations decisions. Based on their findings, the researchers stress the need for strong state, regional, and national policies that encourage and support the steps that homeowners and industry planners can take to help decarbonize today’s building sector.

A two-part modeling approach

To analyze the impacts of electrification of residential heating on costs and emissions in the combined power and gas sectors, a team of MIT experts in building technology, power systems modeling, optimization techniques, and more developed a two-part modeling framework. Team members included Rahman Khorramfar, a senior postdoc in MITEI and the Laboratory for Information and Decision Systems (LIDS); Morgan Santoni-Colvin SM ’23, a former MITEI graduate research assistant, now an associate at Energy and Environmental Economics, Inc.; Saurabh Amin, a professor in the Department of Civil and Environmental Engineering and principal investigator in LIDS; Audun Botterud, a principal research scientist in LIDS; Leslie Norford, a professor in the Department of Architecture; and Dharik Mallapragada, a former MITEI principal research scientist, now an assistant professor at New York University, who led the project. They describe their new methods and findings in a paper published in the journal Cell Reports Sustainability on Feb. 6.

The first model in the framework quantifies how various levels of electrification will change end-use demand for electricity and for natural gas, and the impacts of possible energy-saving measures that homeowners can take to help. “To perform that analysis, we built a ‘bottom-up’ model — meaning that it looks at electricity and gas consumption of individual buildings and then aggregates their consumption to get an overall demand for power and for gas,” explains Khorramfar. By assuming a wide range of building “archetypes” — that is, groupings of buildings with similar physical characteristics and properties — coupled with trends in population growth, the team could explore how demand for electricity and for natural gas would change under each of five assumed electrification pathways: “business as usual” with modest electrification, medium electrification (about 60 percent of homes are electrified), high electrification (about 80 percent of homes make the change), and medium and high electrification with “envelope improvements,” such as sealing up heat leaks and adding insulation.

The second part of the framework consists of a model that takes the demand results from the first model as inputs and “co-optimizes” the overall electricity and natural gas system to minimize annual investment and operating costs while adhering to any constraints, such as limits on emissions or on resource availability. The modeling framework thus enables the researchers to explore the impact of each electrification pathway on the infrastructure and operating costs of the two interacting sectors.

The New England case study: A challenge for electrification

As a case study, the researchers chose New England, a region where the weather is sometimes extremely cold and where burning natural gas to heat houses contributes significantly to overall emissions. “Critics will say that electrification is never going to happen [in New England]. It’s just too expensive,” comments Santoni-Colvin. But he notes that most studies focus on the electricity sector in isolation. The new framework considers the joint operation of the two sectors and then quantifies their respective costs and emissions. “We know that electrification will require large investments in the electricity infrastructure,” says Santoni-Colvin. “But what hasn’t been well quantified in the literature is the savings that we generate on the natural gas side by doing that — so, the system-level savings.”

Using their framework, the MIT team performed model runs aimed at an 80 percent reduction in building-sector emissions relative to 1990 levels — a target consistent with regional policy goals for 2050. The researchers defined parameters including details about building archetypes, the regional electric power system, existing and potential renewable generating systems, battery storage, availability of natural gas, and other key factors describing New England.

They then performed analyses assuming various scenarios with different mixes of home improvements. While most studies assume typical weather, they instead developed 20 projections of annual weather data based on historical weather patterns and adjusted for the effects of climate change through 2050. They then analyzed their five levels of electrification.

Relative to business-as-usual projections, results from the framework showed that high electrification of residential heating could more than double the demand for electricity during peak periods and increase overall electricity demand by close to 60 percent. Assuming that building-envelope improvements are deployed in parallel with electrification reduces the magnitude and weather sensitivity of peak loads and creates overall efficiency gains that reduce the combined demand for electricity plus natural gas for home heating by up to 30 percent relative to the present day. Notably, a combination of high electrification and envelope improvements resulted in the lowest average cost for the overall electric power-natural gas system in 2050.

Lessons learned

Replacing existing natural gas-burning furnaces and boilers with heat pumps reduces overall energy consumption. Santoni-Colvin calls it “something of an intuitive result” that could be expected because heat pumps are “just that much more efficient than old, fossil fuel-burning systems. But even so, we were surprised by the gains.”

Other unexpected results include the importance of homeowners making more traditional energy efficiency improvements, such as adding insulation and sealing air leaks — steps supported by recent rebate policies. Those changes are critical to reducing costs that would otherwise be incurred for upgrading the electricity grid to accommodate the increased demand. “You can’t just go wild dropping heat pumps into everybody’s houses if you’re not also considering other ways to reduce peak loads. So it really requires an ‘all of the above’ approach to get to the most cost-effective outcome,” says Santoni-Colvin.

Testing a range of weather outcomes also provided important insights. Demand for heating fuel is very weather-dependent, yet most studies are based on a limited set of weather data — often a “typical year.” The researchers found that electrification can lead to extended peak electric load events that can last for a few days during cold winters. Accordingly, the researchers conclude that there will be a continuing need for a “firm, dispatchable” source of electricity; that is, a power-generating system that can be relied on to produce power any time it’s needed — unlike solar and wind systems. As examples, they modeled some possible technologies, including power plants fired by a low-carbon fuel or by natural gas equipped with carbon capture equipment. But they point out that there’s no way of knowing what types of firm generators will be available in 2050. It could be a system that’s not yet mature, or perhaps doesn’t even exist today.

In presenting their findings, the researchers note several caveats. For one thing, their analyses don’t include the estimated cost to homeowners of installing heat pumps. While that cost is widely discussed and debated, that issue is outside the scope of their current project.

In addition, the study doesn’t specify what happens to existing natural gas pipelines. “Some homes are going to electrify and get off the gas system and not have to pay for it, leaving other homes with increasing rates because the gas system cost now has to be divided among fewer customers,” says Khorramfar. “That will inevitably raise equity questions that need to be addressed by policymakers.”

Finally, the researchers note that policies are needed to drive residential electrification. Current financial support for installation of heat pumps and steps to make homes more thermally efficient are a good start. But such incentives must be coupled with a new approach to planning energy infrastructure investments. Traditionally, electric power planning and natural gas planning are performed separately. However, to decarbonize residential heating, the two sectors should coordinate when planning future operations and infrastructure needs. Results from the MIT analysis indicate that such cooperation could significantly reduce both emissions and costs for residential heating — a change that would yield a much-needed step toward decarbonizing the buildings sector as a whole.

MIT’s Abdul Latif Jameel Water and Food Systems Lab (J-WAFS) has transformed the landscape of water and food research at MIT, driving faculty engagement and catalyzing new research and innovation in these critical areas. With philanthropic, corporate, and government support, J-WAFS’ strategic approach spans the entire research life cycle, from support for early-stage research to commercialization grants for more advanced projects.

Over the past decade, J-WAFS has invested approximately $25 million in direct research funding to support MIT faculty pursuing transformative research with the potential for significant impact. “Since awarding our first cohort of seed grants in 2015, it’s remarkable to look back and see that over 10 percent of the MIT faculty have benefited from J-WAFS funding,” observes J-WAFS Executive Director Renee J. Robins ’83. “Many of these professors hadn’t worked on water or food challenges before their first J-WAFS grant.” 

By fostering interdisciplinary collaborations and supporting high-risk, high-reward projects, J-WAFS has amplified the capacity of MIT faculty to pursue groundbreaking research that addresses some of the world’s most pressing challenges facing our water and food systems.

Drawing MIT faculty to water and food research

J-WAFS open calls for proposals enable faculty to explore bold ideas and develop impactful approaches to tackling critical water and food system challenges. Professor Patrick Doyle’s work in water purification exemplifies this impact. “Without J-WAFS, I would have never ventured into the field of water purification,” Doyle reflects. While previously focused on pharmaceutical manufacturing and drug delivery, exposure to J-WAFS-funded peers led him to apply his expertise in soft materials to water purification. “Both the funding and the J-WAFS community led me to be deeply engaged in understanding some of the key challenges in water purification and water security,” he explains.

Similarly, Professor Otto Cordero of the Department of Civil and Environmental Engineering (CEE) leveraged J-WAFS funding to pivot his research into aquaculture. Cordero explains that his first J-WAFS seed grant “has been extremely influential for my lab because it allowed me to take a step in a new direction, with no preliminary data in hand.” Cordero’s expertise is in microbial communities. He was previous unfamiliar with aquaculture, but he saw the relevance of microbial communities the health of farmed aquatic organisms.

Supporting early-career faculty

New assistant professors at MIT have particularly benefited from J-WAFS funding and support. J-WAFS has played a transformative role in shaping the careers and research trajectories of many new faculty members by encouraging them to explore novel research areas, and in many instances providing their first MIT research grant.

Professor Ariel Furst reflects on how pivotal J-WAFS’ investment has been in advancing her research. “This was one of the first grants I received after starting at MIT, and it has truly shaped the development of my group’s research program,” Furst explains. With J-WAFS’ backing, her lab has achieved breakthroughs in chemical detection and remediation technologies for water. “The support of J-WAFS has enabled us to develop the platform funded through this work beyond the initial applications to the general detection of environmental contaminants and degradation of those contaminants,” she elaborates. 

Karthish Manthiram, now a professor of chemical engineering and chemistry at Caltech, explains how J-WAFS’ early investment enabled him and other young faculty to pursue ambitious ideas. “J-WAFS took a big risk on us,” Manthiram reflects. His research on breaking the nitrogen triple bond to make ammonia for fertilizer was initially met with skepticism. However, J-WAFS’ seed funding allowed his lab to lay the groundwork for breakthroughs that later attracted significant National Science Foundation (NSF) support. “That early funding from J-WAFS has been pivotal to our long-term success,” he notes. 

These stories underscore the broad impact of J-WAFS’ support for early-career faculty, and its commitment to empowering them to address critical global challenges and innovate boldly.

Fueling follow-on funding 

J-WAFS seed grants enable faculty to explore nascent research areas, but external funding for continued work is usually necessary to achieve the full potential of these novel ideas. “It’s often hard to get funding for early stage or out-of-the-box ideas,” notes J-WAFS Director Professor John H. Lienhard V. “My hope, when I founded J-WAFS in 2014, was that seed grants would allow PIs [principal investigators] to prove out novel ideas so that they would be attractive for follow-on funding. And after 10 years, J-WAFS-funded research projects have brought more than $21 million in subsequent awards to MIT.”

Professor Retsef Levi led a seed study on how agricultural supply chains affect food safety, with a team of faculty spanning the MIT schools Engineering and Science as well as the MIT Sloan School of Management. The team parlayed their seed grant research into a multi-million-dollar follow-on initiative. Levi reflects, “The J-WAFS seed funding allowed us to establish the initial credibility of our team, which was key to our success in obtaining large funding from several other agencies.”

Dave Des Marais was an assistant professor in the Department of CEE when he received his first J-WAFS seed grant. The funding supported his research on how plant growth and physiology are controlled by genes and interact with the environment. The seed grant helped launch his lab’s work addressing enhancing climate change resilience in agricultural systems. The work led to his Faculty Early Career Development (CAREER) Award from the NSF, a prestigious honor for junior faculty members. Now an associate professor, Des Marais’ ongoing project to further investigate the mechanisms and consequences of genomic and environmental interactions is supported by the five-year, $1,490,000 NSF grant. “J-WAFS providing essential funding to get my new research underway,” comments Des Marais.

Stimulating interdisciplinary collaboration

Des Marais’ seed grant was also key to developing new collaborations. He explains, “the J-WAFS grant supported me to develop a collaboration with Professor Caroline Uhler in EECS/IDSS [the Department of Electrical Engineering and Computer Science/Institute for Data, Systems, and Society] that really shaped how I think about framing and testing hypotheses. One of the best things about J-WAFS is facilitating unexpected connections among MIT faculty with diverse yet complementary skill sets.”

Professors A. John Hart of the Department of Mechanical Engineering and Benedetto Marelli of CEE also launched a new interdisciplinary collaboration with J-WAFS funding. They partnered to join expertise in biomaterials, microfabrication, and manufacturing, to create printed silk-based colorimetric sensors that detect food spoilage. “The J-WAFS Seed Grant provided a unique opportunity for multidisciplinary collaboration,” Hart notes.

Professors Stephen Graves in the MIT Sloan School of Management and Bishwapriya Sanyal in the Department of Urban Studies and Planning (DUSP) partnered to pursue new research on agricultural supply chains. With field work in Senegal, their J-WAFS-supported project brought together international development specialists and operations management experts to study how small firms and government agencies influence access to and uptake of irrigation technology by poorer farmers. “We used J-WAFS to spur a collaboration that would have been improbable without this grant,” they explain. Being part of the J-WAFS community also introduced them to researchers in Professor Amos Winter’s lab in the Department of Mechanical Engineering working on irrigation technologies for low-resource settings. DUSP doctoral candidate Mark Brennan notes, “We got to share our understanding of how irrigation markets and irrigation supply chains work in developing economies, and then we got to contrast that with their understanding of how irrigation system models work.”

Timothy Swager, professor of chemistry, and Rohit Karnik, professor of mechanical engineering and J-WAFS associate director, collaborated on a sponsored research project supported by Xylem, Inc. through the J-WAFS Research Affiliate program. The cross-disciplinary research, which targeted the development of ultra-sensitive sensors for toxic PFAS chemicals, was conceived following a series of workshops hosted by J-WAFS. Swager and Karnik were two of the participants, and their involvement led to the collaborative proposal that Xylem funded. “J-WAFS funding allowed us to combine Swager lab’s expertise in sensing with my lab’s expertise in microfluidics to develop a cartridge for field-portable detection of PFAS,” says Karnik. “J-WAFS has enriched my research program in so many ways,” adds Swager, who is now working to commercialize the technology.

Driving global collaboration and impact

J-WAFS has also helped MIT faculty establish and advance international collaboration and impactful global research. By funding and supporting projects that connect MIT researchers with international partners, J-WAFS has not only advanced technological solutions, but also strengthened cross-cultural understanding and engagement.

Professor Matthew Shoulders leads the inaugural J-WAFS Grand Challenge project. In response to the first J-WAFS call for “Grand Challenge” proposals, Shoulders assembled an interdisciplinary team based at MIT to enhance and provide climate resilience to agriculture by improving the most inefficient aspect of photosynthesis, the notoriously-inefficient carbon dioxide-fixing plant enzyme RuBisCO. J-WAFS funded this high-risk/high-reward project following a competitive process that engaged external reviewers through a several rounds of iterative proposal development. The technical feedback to the team led them to researchers with complementary expertise from the Australian National University. “Our collaborative team of biochemists and synthetic biologists, computational biologists, and chemists is deeply integrated with plant biologists and field trial experts, yielding a robust feedback loop for enzyme engineering,” Shoulders says. “Together, this team will be able to make a concerted effort using the most modern, state-of-the-art techniques to engineer crop RuBisCO with an eye to helping make meaningful gains in securing a stable crop supply, hopefully with accompanying improvements in both food and water security.”

Professor Leon Glicksman and Research Engineer Eric Verploegen’s team designed a low-cost cooling chamber to preserve fruits and vegetables harvested by smallholder farmers with no access to cold chain storage. J-WAFS’ guidance motivated the team to prioritize practical considerations informed by local collaborators, ensuring market competitiveness. “As our new idea for a forced-air evaporative cooling chamber was taking shape, we continually checked that our solution was evolving in a direction that would be competitive in terms of cost, performance, and usability to existing commercial alternatives,” explains Verploegen, who is currently an MIT D-Lab affiliate. Following the team’s initial seed grant, the team secured a J-WAFS Solutions commercialization grant, which Verploegen say “further motivated us to establish partnerships with local organizations capable of commercializing the technology earlier in the project than we might have done otherwise.” The team has since shared an open-source design as part of its commercialization strategy to maximize accessibility and impact.

Bringing corporate sponsored research opportunities to MIT faculty

J-WAFS also plays a role in driving private partnerships, enabling collaborations that bridge industry and academia. Through its Research Affiliate Program, for example, J-WAFS provides opportunities for faculty to collaborate with industry on sponsored research, helping to convert scientific discoveries into licensable intellectual property (IP) that companies can turn into commercial products and services.

J-WAFS introduced professor of mechanical engineering Alex Slocum to a challenge presented by its research affiliate company, Xylem: how to design a more energy-efficient pump for fluctuating flows. With centrifugal pumps consuming an estimated 6 percent of U.S. electricity annually, Slocum and his then-graduate student Hilary Johnson SM '18, PhD '22 developed an innovative variable volute mechanism that reduces energy usage. “Xylem envisions this as the first in a new category of adaptive pump geometry,” comments Johnson. The research produced a pump prototype and related IP that Xylem is working on commercializing. Johnson notes that these outcomes “would not have been possible without J-WAFS support and facilitation of the Xylem industry partnership.” Slocum adds, “J-WAFS enabled Hilary to begin her work on pumps, and Xylem sponsored the research to bring her to this point … where she has an opportunity to do far more than the original project called for.”

Swager speaks highly of the impact of corporate research sponsorship through J-WAFS on his research and technology translation efforts. His PFAS project with Karnik described above was also supported by Xylem. “Xylem was an excellent sponsor of our research. Their engagement and feedback were instrumental in advancing our PFAS detection technology, now on the path to commercialization,” Swager says.

Looking forward

What J-WAFS has accomplished is more than a collection of research projects; a decade of impact demonstrates how J-WAFS’ approach has been transformative for many MIT faculty members. As Professor Mathias Kolle puts it, his engagement with J-WAFS “had a significant influence on how we think about our research and its broader impacts.” He adds that it “opened my eyes to the challenges in the field of water and food systems and the many different creative ideas that are explored by MIT.” 

This thriving ecosystem of innovation, collaboration, and academic growth around water and food research has not only helped faculty build interdisciplinary and international partnerships, but has also led to the commercialization of transformative technologies with real-world applications. C. Cem Taşan, the POSCO Associate Professor of Metallurgy who is leading a J-WAFS Solutions commercialization team that is about to launch a startup company, sums it up by noting, “Without J-WAFS, we wouldn’t be here at all.”  

As J-WAFS looks to the future, its continued commitment — supported by the generosity of its donors and partners — builds on a decade of success enabling MIT faculty to advance water and food research that addresses some of the world’s most pressing challenges.

While early language models could only process text, contemporary large language models now perform highly diverse tasks on different types of data. For instance, LLMs can understand many languages, generate computer code, solve math problems, or answer questions about images and audio.   

MIT researchers probed the inner workings of LLMs to better understand how they process such assorted data, and found evidence that they share some similarities with the human brain.

Neuroscientists believe the human brain has a “semantic hub” in the anterior temporal lobe that integrates semantic information from various modalities, like visual data and tactile inputs. This semantic hub is connected to modality-specific “spokes” that route information to the hub. The MIT researchers found that LLMs use a similar mechanism by abstractly processing data from diverse modalities in a central, generalized way. For instance, a model that has English as its dominant language would rely on English as a central medium to process inputs in Japanese or reason about arithmetic, computer code, etc. Furthermore, the researchers demonstrate that they can intervene in a model’s semantic hub by using text in the model’s dominant language to change its outputs, even when the model is processing data in other languages.

These findings could help scientists train future LLMs that are better able to handle diverse data.

“LLMs are big black boxes. They have achieved very impressive performance, but we have very little knowledge about their internal working mechanisms. I hope this can be an early step to better understand how they work so we can improve upon them and better control them when needed,” says Zhaofeng Wu, an electrical engineering and computer science (EECS) graduate student and lead author of a paper on this research.

His co-authors include Xinyan Velocity Yu, a graduate student at the University of Southern California (USC); Dani Yogatama, an associate professor at USC; Jiasen Lu, a research scientist at Apple; and senior author Yoon Kim, an assistant professor of EECS at MIT and a member of the Computer Science and Artificial Intelligence Laboratory (CSAIL). The research will be presented at the International Conference on Learning Representations.

Integrating diverse data

The researchers based the new study upon prior work which hinted that English-centric LLMs use English to perform reasoning processes on various languages.

Wu and his collaborators expanded this idea, launching an in-depth study into the mechanisms LLMs use to process diverse data.

An LLM, which is composed of many interconnected layers, splits input text into words or sub-words called tokens. The model assigns a representation to each token, which enables it to explore the relationships between tokens and generate the next word in a sequence. In the case of images or audio, these tokens correspond to particular regions of an image or sections of an audio clip.

The researchers found that the model’s initial layers process data in its specific language or modality, like the modality-specific spokes in the human brain. Then, the LLM converts tokens into modality-agnostic representations as it reasons about them throughout its internal layers, akin to how the brain’s semantic hub integrates diverse information.

The model assigns similar representations to inputs with similar meanings, despite their data type, including images, audio, computer code, and arithmetic problems. Even though an image and its text caption are distinct data types, because they share the same meaning, the LLM would assign them similar representations.

For instance, an English-dominant LLM “thinks” about a Chinese-text input in English before generating an output in Chinese. The model has a similar reasoning tendency for non-text inputs like computer code, math problems, or even multimodal data.

To test this hypothesis, the researchers passed a pair of sentences with the same meaning but written in two different languages through the model. They measured how similar the model’s representations were for each sentence.

Then they conducted a second set of experiments where they fed an English-dominant model text in a different language, like Chinese, and measured how similar its internal representation was to English versus Chinese. The researchers conducted similar experiments for other data types.

They consistently found that the model’s representations were similar for sentences with similar meanings. In addition, across many data types, the tokens the model processed in its internal layers were more like English-centric tokens than the input data type.

“A lot of these input data types seem extremely different from language, so we were very surprised that we can probe out English-tokens when the model processes, for example, mathematic or coding expressions,” Wu says.

Leveraging the semantic hub

The researchers think LLMs may learn this semantic hub strategy during training because it is an economical way to process varied data.

“There are thousands of languages out there, but a lot of the knowledge is shared, like commonsense knowledge or factual knowledge. The model doesn’t need to duplicate that knowledge across languages,” Wu says.

The researchers also tried intervening in the model’s internal layers using English text when it was processing other languages. They found that they could predictably change the model outputs, even though those outputs were in other languages.

Scientists could leverage this phenomenon to encourage the model to share as much information as possible across diverse data types, potentially boosting efficiency.

But on the other hand, there could be concepts or knowledge that are not translatable across languages or data types, like culturally specific knowledge. Scientists might want LLMs to have some language-specific processing mechanisms in those cases.

“How do you maximally share whenever possible but also allow languages to have some language-specific processing mechanisms? That could be explored in future work on model architectures,” Wu says.

In addition, researchers could use these insights to improve multilingual models. Often, an English-dominant model that learns to speak another language will lose some of its accuracy in English. A better understanding of an LLM’s semantic hub could help researchers prevent this language interference, he says.

“Understanding how language models process inputs across languages and modalities is a key question in artificial intelligence. This paper makes an interesting connection to neuroscience and shows that the proposed ‘semantic hub hypothesis’ holds in modern language models, where semantically similar representations of different data types are created in the model’s intermediate layers,” says Mor Geva Pipek, an assistant professor in the School of Computer Science at Tel Aviv University, who was not involved with this work. “The hypothesis and experiments nicely tie and extend findings from previous works and could be influential for future research on creating better multimodal models and studying links between them and brain function and cognition in humans.”

This research is funded, in part, by the MIT-IBM Watson AI Lab.

Creating and sustaining fusion reactions — essentially recreating star-like conditions on Earth — is extremely difficult, and Nathan Howard PhD ’12, a principal research scientist at the MIT Plasma Science and Fusion Center (PSFC), thinks it’s one of the most fascinating scientific challenges of our time. “Both the science and the overall promise of fusion as a clean energy source are really interesting. That motivated me to come to grad school [at MIT] and work at the PSFC,” he says.

Howard is member of the Magnetic Fusion Experiments Integrated Modeling (MFE-IM) group at the PSFC. Along with MFE-IM group leader Pablo Rodriguez-Fernandez, Howard and the team use simulations and machine learning to predict how plasma will behave in a fusion device. MFE-IM and Howard’s research aims to forecast a given technology or configuration’s performance before it’s piloted in an actual fusion environment, allowing for smarter design choices. To ensure their accuracy, these models are continuously validated using data from previous experiments, keeping their simulations grounded in reality.

In a recent open-access paper titled “Prediction of Performance and Turbulence in ITER Burning Plasmas via Nonlinear Gyrokinetic Profile Prediction,” published in the January issue of Nuclear Fusion, Howard explains how he used high-resolution simulations of the swirling structures present in plasma, called turbulence, to confirm that the world’s largest experimental fusion device, currently under construction in Southern France, will perform as expected when switched on. He also demonstrates how a different operating setup could produce nearly the same amount of energy output but with less energy input, a discovery that could positively affect the efficiency of fusion devices in general.

The biggest and best of what’s never been built

Forty years ago, the United States and six other member nations came together to build ITER (Latin for “the way”), a fusion device that, once operational, would yield 500 megawatts of fusion power, and a plasma able to generate 10 times more energy than it absorbs from external heating. The plasma setup designed to achieve these goals — the most ambitious of any fusion experiment — is called the ITER baseline scenario, and as fusion science and plasma physics have progressed, ways to achieve this plasma have been refined using increasingly more powerful simulations like the modeling framework Howard used.

In his work to verify the baseline scenario, Howard used CGYRO, a computer code developed by Howard’s collaborators at General Atomics. CGYRO applies a complex plasma physics model to a set of defined fusion operating conditions. Although it is time-intensive, CGYRO generates very detailed simulations on how plasma behaves at different locations within a fusion device.

The comprehensive CGYRO simulations were then run through the PORTALS framework, a collection of tools originally developed at MIT by Rodriguez-Fernandez. “PORTALS takes the high-fidelity [CGYRO] runs and uses machine learning to build a quick model called a ‘surrogate’ that can mimic the results of the more complex runs, but much faster,” Rodriguez-Fernandez explains. “Only high-fidelity modeling tools like PORTALS give us a glimpse into the plasma core before it even forms. This predict-first approach allows us to create more efficient plasmas in a device like ITER.”

After the first pass, the surrogates’ accuracy was checked against the high-fidelity runs, and if a surrogate wasn’t producing results in line with CGYRO’s, PORTALS was run again to refine the surrogate until it better mimicked CGYRO’s results. “The nice thing is, once you have built a well-trained [surrogate] model, you can use it to predict conditions that are different, with a very much reduced need for the full complex runs.” Once they were fully trained, the surrogates were used to explore how different combinations of inputs might affect ITER’s predicted performance and how it achieved the baseline scenario. Notably, the surrogate runs took a fraction of the time, and they could be used in conjunction with CGYRO to give it a boost and produce detailed results more quickly.

“Just dropped in to see what condition my condition was in”

Howard’s work with CGYRO, PORTALS, and surrogates examined a specific combination of operating conditions that had been predicted to achieve the baseline scenario. Those conditions included the magnetic field used, the methods used to control plasma shape, the external heating applied, and many other variables. Using 14 iterations of CGYRO, Howard was able to confirm that the current baseline scenario configuration could achieve 10 times more power output than input into the plasma. Howard says of the results, “The modeling we performed is maybe the highest fidelity possible at this time, and almost certainly the highest fidelity published.”

The 14 iterations of CGYRO used to confirm the plasma performance included running PORTALS to build surrogate models for the input parameters and then tying the surrogates to CGYRO to work more efficiently. It only took three additional iterations of CGYRO to explore an alternate scenario that predicted ITER could produce almost the same amount of energy with about half the input power. The surrogate-enhanced CGYRO model revealed that the temperature of the plasma core — and thus the fusion reactions — wasn’t overly affected by less power input; less power input equals more efficient operation. Howard’s results are also a reminder that there may be other ways to improve ITER’s performance; they just haven’t been discovered yet.

Howard reflects, “The fact that we can use the results of this modeling to influence the planning of experiments like ITER is exciting. For years, I’ve been saying that this was the goal of our research, and now that we actually do it — it’s an amazing arc, and really fulfilling.” 

Within the animal kingdom, mussels are masters of underwater adhesion. The marine molluscs cluster atop rocks and along the bottoms of ships, and hold fast against the ocean’s waves thanks to a gluey plaque they secrete through their foot. These tenacious adhesive structures have prompted scientists in recent years to design similar bioinspired, waterproof adhesives.

Now engineers from MIT and Freie Universität Berlin have developed a new type of glue that combines the waterproof stickiness of the mussels’ plaques with the germ-proof properties of another natural material: mucus.

Every surface in our bodies not covered in skin is lined with a protective layer of mucus — a slimy network of proteins that acts as a physical barrier against bacteria and other infectious agents. In their new work, the engineers combined sticky, mussel-inspired polymers with mucus-derived proteins, or mucins, to form a gel that strongly adheres to surfaces.

The new mucus-derived glue prevented the buildup of bacteria while keeping its sticky hold, even on wet surfaces. The researchers envision that once the glue’s properties are optimized, it could be applied as a liquid by injection or spray, which would then solidify into a sticky gel. The material might be used to coat medical implants, for example, to prevent infection and bacteria buildup.

The team’s new glue-making approach could also be adjusted to incorporate other natural materials, such as keratin — a fibrous substance found in feathers and hair, with certain chemical features resembling those of mucus.

“The applications of our materials design approach will depend on the specific precursor materials,” says George Degen, a postdoc in MIT’s Department of Mechanical Engineering. “For example, mucus-derived or mucus-inspired materials might be used as multifunctional biomedical adhesives that also prevent infections. Alternatively, applying our approach to keratin might enable development of sustainable packaging materials.”

A paper detailing the team’s results appears this week in the Proceedings of the National Academy of Sciences. Degen’s MIT co-authors include Corey Stevens, Gerardo Cárcamo-Oyarce, Jake Song, Katharina Ribbeck, and Gareth McKinley, along with Raju Bej, Peng Tang, and Rainer Haag of Freie Universität Berlin.

A sticky combination

Before coming to MIT, Degen was a graduate student at the University of California at Santa Barbara, where he worked in a research group that studied the adhesive mechanisms of mussels.

“Mussels are able to deposit materials that adhere to wet surfaces in seconds to minutes,” Degen says. “These natural materials do better than existing commercialized adhesives, specifically at sticking to wet and underwater surfaces, which has been a longstanding technical challenge.”

To stick to a rock or a ship, mussels secrete a protein-rich fluid. Chemical bonds, or cross-links, act as connection points between proteins, enabling the secreted substance to simultaneously solidify into a gel and stick to a wet surface.

As it happens, similar cross-linking features are found in mucin — a large protein that is the primary non-water component of mucus. When Degen came to MIT, he worked with both McKinley, a professor of mechanical engineering and an expert in materials science and fluid flow, and Katharina Ribbeck, a professor of biological engineering and a leader in the study of mucus, to develop a cross-linking glue that would combine the adhesive qualities of mussel plaques with the bacteria-blocking properties of mucus.

Mixing links

The MIT researchers teamed up with Haag and colleagues in Berlin who specialize in synthesizing bioinspired materials. Haag and Ribbeck are members of a collaborative research group that develops dynamic hydrogels for biointerfaces. Haag’s group has made mussel-like adhesives, as well as mucus-inspired liquids by producing microscopic, fiber-like polymers that are similar in structure to the natural mucin proteins.

For their new work, the researchers focused on a chemical motif that appears in mussel adhesives: a bond between two chemical groups known as “catechols” and “thiols.” In the mussel’s natural glue, or plaque, these groups combine to form catechol–thiol cross-links that contribute to the cohesive strength of the plaque. Catechols also enhance a mussel’s adhesion by binding to surfaces such as rocks and ship hulls.

Interestingly, thiol groups are also prevalent in mucin proteins. Degen wondered whether mussel-inspired polymers could link with mucin thiols, enabling the mucins to quickly turn from a liquid to a sticky gel.

To test this idea, he combined solutions of natural mucin proteins with synthetic mussel-inspired polymers and observed how the resulting mixture solidified and stuck to surfaces over time.

“It’s like a two-part epoxy. You combine two liquids together, and chemistry starts to occur so that the liquid solifidies while the substance is simultaneously glueing itself to the surface,” Degen says. 

“Depending on how much cross-linking you have, we can control the speed at which the liquids gelate and adhere,” Haag adds. “We can do this all on wet surfaces, at room temperature, and under very mild conditions. This is what is quite unique.”

The team deposited a range of compositions between two surfaces and found that the resulting adhesive held the surfaces together, with forces comparable to the commercial medical adhesives used for bonding tissue. The researchers also tested the adhesive’s bacteria-blocking properties by depositing the gel onto glass surfaces and incubating them with bacteria overnight.

“We found if we had a bare glass surface without our coating, the bacteria formed a thick biofilm, whereas with our coating, biofilms were largely prevented,” Degen notes.

The team says that with a bit of tuning, they can further improve the adhesive’s hold. Then, the material could be a strong and protective alternative to existing medical adhesives.

“We are excited to have established a biomaterials design platform that gives us these desirable properties of gelation and adhesion, and as a starting point we’ve demonstrated some key biomedical applications,” Degen says. “We are now ready to expand into different synthetic and natural systems and target different applications.”

This research was funded, in part, by the U.S. National Institutes of Health, the U.S. National Science Foundation, and the U.S. Army Research Office.

Proteins are the workhorses that keep our cells running, and there are many thousands of types of proteins in our cells, each performing a specialized function. Researchers have long known that the structure of a protein determines what it can do. More recently, researchers are coming to appreciate that a protein’s localization is also critical for its function. Cells are full of compartments that help to organize their many denizens. Along with the well-known organelles that adorn the pages of biology textbooks, these spaces also include a variety of dynamic, membrane-less compartments that concentrate certain molecules together to perform shared functions. Knowing where a given protein localizes, and who it co-localizes with, can therefore be useful for better understanding that protein and its role in the healthy or diseased cell, but researchers have lacked a systematic way to predict this information.

Meanwhile, protein structure has been studied for over half-a-century, culminating in the artificial intelligence tool AlphaFold, which can predict protein structure from a protein’s amino acid code, the linear string of building blocks within it that folds to create its structure. AlphaFold and models like it have become widely used tools in research.

Proteins also contain regions of amino acids that do not fold into a fixed structure, but are instead important for helping proteins join dynamic compartments in the cell. MIT Professor Richard Young and colleagues wondered whether the code in those regions could be used to predict protein localization in the same way that other regions are used to predict structure. Other researchers have discovered some protein sequences that code for protein localization, and some have begun developing predictive models for protein localization. However, researchers did not know whether a protein’s localization to any dynamic compartment could be predicted based on its sequence, nor did they have a comparable tool to AlphaFold for predicting localization. 

Now, Young, also member of the Whitehead Institute for Biological Research; Young lab postdoc Henry Kilgore; Regina Barzilay, the School of Engineering Distinguished Professor for AI and Health in MIT's Department of Electrical Engineering and Computer Science and principal investigator in the Computer Science and Artificial Intelligence Laboratory (CSAIL); and colleagues have built such a model, which they call ProtGPS. In a paper published on Feb. 6 in the journal Science, with first authors Kilgore and Barzilay lab graduate students Itamar Chinn, Peter Mikhael, and Ilan Mitnikov, the cross-disciplinary team debuts their model. The researchers show that ProtGPS can predict to which of 12 known types of compartments a protein will localize, as well as whether a disease-associated mutation will change that localization. Additionally, the research team developed a generative algorithm that can design novel proteins to localize to specific compartments.

“My hope is that this is a first step towards a powerful platform that enables people studying proteins to do their research,” Young says, “and that it helps us understand how humans develop into the complex organisms that they are, how mutations disrupt those natural processes, and how to generate therapeutic hypotheses and design drugs to treat dysfunction in a cell.”

The researchers also validated many of the model’s predictions with experimental tests in cells.

“It really excited me to be able to go from computational design all the way to trying these things in the lab,” Barzilay says. “There are a lot of exciting papers in this area of AI, but 99.9 percent of those never get tested in real systems. Thanks to our collaboration with the Young lab, we were able to test, and really learn how well our algorithm is doing.”

Developing the model

The researchers trained and tested ProtGPS on two batches of proteins with known localizations. They found that it could correctly predict where proteins end up with high accuracy. The researchers also tested how well ProtGPS could predict changes in protein localization based on disease-associated mutations within a protein. Many mutations — changes to the sequence for a gene and its corresponding protein — have been found to contribute to or cause disease based on association studies, but the ways in which the mutations lead to disease symptoms remain unknown.

Figuring out the mechanism for how a mutation contributes to disease is important because then researchers can develop therapies to fix that mechanism, preventing or treating the disease. Young and colleagues suspected that many disease-associated mutations might contribute to disease by changing protein localization. For example, a mutation could make a protein unable to join a compartment containing essential partners.

They tested this hypothesis by feeding ProtGOS more than 200,000 proteins with disease-associated mutations, and then asking it to both predict where those mutated proteins would localize and measure how much its prediction changed for a given protein from the normal to the mutated version. A large shift in the prediction indicates a likely change in localization.

The researchers found many cases in which a disease-associated mutation appeared to change a protein’s localization. They tested 20 examples in cells, using fluorescence to compare where in the cell a normal protein and the mutated version of it ended up. The experiments confirmed ProtGPS’s predictions. Altogether, the findings support the researchers’ suspicion that mis-localization may be an underappreciated mechanism of disease, and demonstrate the value of ProtGPS as a tool for understanding disease and identifying new therapeutic avenues.

“The cell is such a complicated system, with so many components and complex networks of interactions,” Mitnikov says. “It’s super interesting to think that with this approach, we can perturb the system, see the outcome of that, and so drive discovery of mechanisms in the cell, or even develop therapeutics based on that.”

The researchers hope that others begin using ProtGPS in the same way that they use predictive structural models like AlphaFold, advancing various projects on protein function, dysfunction, and disease.

Moving beyond prediction to novel generation

The researchers were excited about the possible uses of their prediction model, but they also wanted their model to go beyond predicting localizations of existing proteins, and allow them to design completely new proteins. The goal was for the model to make up entirely new amino acid sequences that, when formed in a cell, would localize to a desired location. Generating a novel protein that can actually accomplish a function — in this case, the function of localizing to a specific cellular compartment — is incredibly difficult. In order to improve their model’s chances of success, the researchers constrained their algorithm to only design proteins like those found in nature. This is an approach commonly used in drug design, for logical reasons; nature has had billions of years to figure out which protein sequences work well and which do not.

Because of the collaboration with the Young lab, the machine learning team was able to test whether their protein generator worked. The model had good results. In one round, it generated 10 proteins intended to localize to the nucleolus. When the researchers tested these proteins in the cell, they found that four of them strongly localized to the nucleolus, and others may have had slight biases toward that location as well.

“The collaboration between our labs has been so generative for all of us,” Mikhael says. “We’ve learned how to speak each other’s languages, in our case learned a lot about how cells work, and by having the chance to experimentally test our model, we’ve been able to figure out what we need to do to actually make the model work, and then make it work better.”

Being able to generate functional proteins in this way could improve researchers’ ability to develop therapies. For example, if a drug must interact with a target that localizes within a certain compartment, then researchers could use this model to design a drug to also localize there. This should make the drug more effective and decrease side effects, since the drug will spend more time engaging with its target and less time interacting with other molecules, causing off-target effects.

The machine learning team members are enthused about the prospect of using what they have learned from this collaboration to design novel proteins with other functions beyond localization, which would expand the possibilities for therapeutic design and other applications.

“A lot of papers show they can design a protein that can be expressed in a cell, but not that the protein has a particular function,” Chinn says. “We actually had functional protein design, and a relatively huge success rate compared to other generative models. That’s really exciting to us, and something we would like to build on.”

All of the researchers involved see ProtGPS as an exciting beginning. They anticipate that their tool will be used to learn more about the roles of localization in protein function and mis-localization in disease. In addition, they are interested in expanding the model’s localization predictions to include more types of compartments, testing more therapeutic hypotheses, and designing increasingly functional proteins for therapies or other applications.

“Now that we know that this protein code for localization exists, and that machine learning models can make sense of that code and even create functional proteins using its logic, that opens up the door for so many potential studies and applications,” Kilgore says.

In the future, autonomous drones could be used to shuttle inventory between large warehouses. A drone might fly into a semi-dark structure the size of several football fields, zipping along hundreds of identical aisles before docking at the precise spot where its shipment is needed.

Most of today’s drones would likely struggle to complete this task, since drones typically navigate outdoors using GPS, which doesn’t work in indoor environments. For indoor navigation, some drones employ computer vision or lidar, but both techniques are unreliable in dark environments or rooms with plain walls or repetitive features.

MIT researchers have introduced a new approach that enables a drone to self-localize, or determine its position, in indoor, dark, and low-visibility environments. Self-localization is a key step in autonomous navigation.

The researchers developed a system called MiFly, in which a drone uses radio frequency (RF) waves, reflected by a single tag placed in its environment, to autonomously self-localize.

Because MiFly enables self-localization with only one small tag, which could be affixed to a wall like a sticker, it would be cheaper and easier to implement than systems that require multiple tags. In addition, since the MiFly tag reflects signals sent by the drone, rather than generating its own signal, it can be operated with very low power.

Two off-the-shelf radars mounted on the drone enable it to localize in relation to the tag. Those measurements are fused with data from the drone’s onboard computer, which enables it to estimate its trajectory.

The researchers conducted hundreds of flight experiments with real drones in indoor environments, and found that MiFly consistently localized the drone to within fewer than 7 centimeters.

“As our understanding of perception and computing improves, we often forget about signals that are beyond the visible spectrum. Here, we’ve looked beyond GPS and computer vision to millimeter waves, and by doing so, we’ve opened up new capabilities for drones in indoor environments that were not possible before,” says Fadel Adib, associate professor in the Department of Electrical Engineering and Computer Science, director of the Signal Kinetics group in the MIT Media Lab, and senior author of a paper on MiFly.

Adib is joined on the paper by co-lead authors and research assistants Maisy Lam and Laura Dodds; Aline Eid, a former postdoc who is now an assistant professor at the University of Michigan; and Jimmy Hester, CTO and co-founder of Atheraxon, Inc. The research will be presented at the IEEE Conference on Computer Communications.

Backscattered signals

To enable drones to self-localize within dark, indoor environments, the researchers decided to utilize millimeter wave signals. Millimeter waves, which are commonly used in modern radars and 5G communication systems, work in the dark and can travel through everyday materials like cardboard, plastic, and interior walls.

They set out to create a system that could work with just one tag, so it would be cheaper and easier to implement in commercial environments. To ensure the device remained low power, they designed a backscatter tag that reflects millimeter wave signals sent by a drone’s onboard radar. The drone uses those reflections to self-localize.

But the drone’s radar would receive signals reflected from all over the environment, not just the tag. The researchers surmounted this challenge by employing a technique called modulation. They configured the tag to add a small frequency to the signal it scatters back to the drone.

“Now, the reflections from the surrounding environment come back at one frequency, but the reflections from the tag come back at a different frequency. This allows us to separate the responses and just look at the response from the tag,” Dodds says.

However, with just one tag and one radar, the researchers could only calculate distance measurements. They needed multiple signals to compute the drone’s location.

Rather than using more tags, they added a second radar to the drone, mounting one horizontally and one vertically. The horizontal radar has a horizontal polarization, which means it sends signals horizontally, while the vertical radar would have a vertical polarization.

They incorporated polarization into the tag’s antennas so it could isolate the separate signals sent by each radar.

“Polarized sunglasses receive a certain polarization of light and block out other polarizations. We applied the same concept to millimeter waves,” Lam explains.

In addition, they applied different modulation frequencies to the vertical and horizontal signals, further reducing interference.

Precise location estimation

This dual-polarization and dual-modulation architecture gives the drone’s spatial location. But drones also move at an angle and rotate, so to enable a drone to navigate, it must estimate its position in space with respect to six degrees of freedom — with trajectory data including pitch, yaw, and roll in addition to the usual forward/backward, left/right, and up/down.

“The drone rotation adds a lot of ambiguity to the millimeter wave estimates. This is a big problem because drones rotate quite a bit as they are flying,” Dodds says.

They overcame these challenges by utilizing the drone’s onboard inertial measurement unit, a sensor that measures acceleration as well as changes in altitude and attitude. By fusing this information with the millimeter wave measurements reflected by the tag, they enable MiFly to estimate the full six-degree-of-freedom pose of the drone in only a few milliseconds.

They tested a MiFly-equipped drone in several indoor environments, including their lab, the flight space at MIT, and the dim tunnels beneath the campus buildings. The system achieved high accuracy consistently across all environments, localizing the drone to within 7 centimeters in many experiments.

In addition, the system was nearly as accurate in situations where the tag was blocked from the drone’s view. They achieved reliable localization estimates up to 6 meters from the tag.

That distance could be extended in the future with the use of additional hardware, such as high-power amplifiers, or by improving the radar and antenna design. The researchers also plan to conduct further research by incorporating MiFly into an autonomous navigation system. This could enable a drone to decide where to fly and execute a flight path using millimeter wave technology.

“The infrastructure and localization algorithms we build up for this work are a strong foundation to go on and make them more robust to enable diverse commercial applications,” Lam says.

This research is funded, in part, by the National Science Foundation and the MIT Media Lab.

The core of a massive cluster of galaxies appears to be pumping out far more stars than it should. Now researchers at MIT and elsewhere have discovered a key ingredient within the cluster that explains the core’s prolific starburst.

In a new study published in Nature, the scientists report using NASA’s James Webb Space Telescope (JWST) to observe the Phoenix cluster — a sprawling collection of gravitationally bound galaxies that circle a central massive galaxy some 5.8 billion light years from Earth. The cluster is the largest of its kind that scientists have so far observed. For its size and estimated age, the Phoenix should be what astronomers call “red and dead” — long done with any star formation that is characteristic of younger galaxies.

But astronomers previously discovered that the core of the Phoenix cluster appeared surprisingly bright, and the central galaxy seemed to be churning out stars at an extremely vigorous rate. The observations raised a mystery: How was the Phoenix fueling such rapid star formation?

In younger galaxies, the “fuel” for forging stars is in the form of extremely cold and dense clouds of interstellar gas. For the much older Phoenix cluster, it was unclear whether the central galaxy could undergo the extreme cooling of gas that would be required to explain its stellar production, or whether cold gas migrated in from other, younger galaxies.

Now, the MIT team has gained a much clearer view of the cluster’s core, using JWST’s far-reaching, infrared-measuring capabilities. For the first time, they have been able to map regions within the core where there are pockets of “warm” gas. Astronomers have previously seen hints of both very hot gas, and very cold gas, but nothing in between.

The detection of warm gas confirms that the Phoenix cluster is actively cooling and able to generate a huge amount of stellar fuel on its own.

“For the first time we have a complete picture of the hot-to-warm-to-cold phase in star formation, which has really never been observed in any galaxy,” says study lead author Michael Reefe, a physics graduate student in MIT’s Kavli Institute for Astrophysics and Space Research. “There is a halo of this intermediate gas everywhere that we can see.”

“The question now is, why this system?” adds co-author Michael McDonald, associate professor of physics at MIT. “This huge starburst could be something every cluster goes through at some point, but we’re only seeing it happen currently in one cluster. The other possibility is that there’s something divergent about this system, and the Phoenix went down a path that other systems don’t go. That would be interesting to explore.”

Hot and cold

The Phoenix cluster was first spotted in 2010 by astronomers using the South Pole Telescope in Antarctica. The cluster comprises about 1,000 galaxies and lies in the constellation Phoenix, after which it is named. Two years later, McDonald led an effort to focus in on Phoenix using multiple telescopes, and discovered that the cluster’s central galaxy was extremely bright. The unexpected luminosity was due to a firehose of star formation. He and his colleagues estimated that this central galaxy was turning out stars at a staggering rate of about 1,000 per year.

“Previous to the Phoenix, the most star-forming galaxy cluster in the universe had about 100 stars per year, and even that was an outlier. The typical number is one-ish,” McDonald says. “The Phoenix is really offset from the rest of the population.”

Since that discovery, scientists have checked in on the cluster from time to time for clues to explain the abnormally high stellar production. They have observed pockets of both ultrahot gas, of about 1 million degrees Fahrenheit, and regions of extremely cold gas, of 10 kelvins, or 10 degrees above absolute zero.

The presence of very hot gas is no surprise: Most massive galaxies, young and old, host black holes at their cores that emit jets of extremely energetic particles that can continually heat up the galaxy’s gas and dust throughout a galaxy’s lifetime. Only in a galaxy’s early stages does some of this million-degree gas cool dramatically to ultracold temperatures that can then form stars. For the Phoenix cluster’s central galaxy, which should be well past the stage of extreme cooling, the presence of ultracold gas presented a puzzle.

“The question has been: Where did this cold gas come from?” McDonald says. “It’s not a given that hot gas will ever cool, because there could be black hole or supernova feedback. So, there are a few viable options, the simplest being that this cold gas was flung into the center from other nearby galaxies. The other is that this gas somehow is directly cooling from the hot gas in the core.”

Neon signs

For their new study, the researchers worked under a key assumption: If the Phoenix cluster’s cold, star-forming gas is coming from within the central galaxy, rather than from the surrounding galaxies, the central galaxy should have not only pockets of hot and cold gas, but also gas that’s in a “warm” in-between phase. Detecting such intermediate gas would be like catching the gas in the midst of extreme cooling, serving as proof that the core of the cluster was indeed the source of the cold stellar fuel.

Following this reasoning, the team sought to detect any warm gas within the Phoenix core. They looked for gas that was somewhere between 10 kelvins and 1 million kelvins. To search for this Goldilocks gas in a system that is 5.8 billion light years away, the researchers looked to JWST, which is capable of observing farther and more clearly than any observatory to date.

The team used the Medium-Resolution Spectrometer on JWST’s Mid-Infrared Instrument (MIRI), which enables scientists to map light in the infrared spectrum. In July of 2023, the team focused the instrument on the Phoenix core and collected 12 hours’ worth of infrared images. They looked for a specific wavelength that is emitted when gas — specifically neon gas — undergoes a certain loss of ions. This transition occurs at around 300,000 kelvins, or 540,000 degrees Fahrenheit — a temperature that happens to be within the “warm” range that the researchers looked to detect and map. The team analyzed the images and mapped the locations where warm gas was observed within the central galaxy.

“This 300,000-degree gas is like a neon sign that’s glowing in a specific wavelength of light, and we could see clumps and filaments of it throughout our entire field of view,” Reefe says. “You could see it everywhere.”

Based on the extent of warm gas in the core, the team estimates that the central galaxy is undergoing a huge degree of extreme cooling and is generating an amount of ultracold gas each year that is equal to the mass of about 20,000 suns. With that kind of stellar fuel supply, the team says it’s very likely that the central galaxy is indeed generating its own starburst, rather than using fuel from surrounding galaxies.

“I think we understand pretty completely what is going on, in terms of what is generating all these stars,” McDonald says. “We don’t understand why. But this new work has opened a new way to observe these systems and understand them better.”

This work was funded, in part, by NASA.

An electrospray engine applies an electric field to a conductive liquid, generating a high-speed jet of tiny droplets that can propel a spacecraft. These miniature engines are ideal for small satellites called CubeSats that are often used in academic research.

Since electrospray engines utilize propellant more efficiently than the powerful, chemical rockets used on the launchpad, they are better suited for precise, in-orbit maneuvers. The thrust generated by an electrospray emitter is tiny, so electrospray engines typically use an array of emitters that are uniformly operated in parallel.

However, these multiplexed electrospray thrusters are typically made via expensive and time-consuming semiconductor cleanroom fabrication, which limits who can manufacture them and how the devices can be applied.

To help break down barriers to space research, MIT engineers have demonstrated the first fully 3D-printed, droplet-emitting electrospray engine. Their device, which can be produced rapidly and for a fraction of the cost of traditional thrusters, uses commercially accessible 3D printing materials and techniques. The devices could even be fully made in orbit, as 3D printing is compatible with in-space manufacturing.

By developing a modular process that combines two 3D printing methods, the researchers overcame the challenges involved in fabricating a complex device comprised of macroscale and microscale components that must work together seamlessly.

Their proof-of-concept thruster comprises 32 electrospray emitters that operate together, generating a stable and uniform flow of propellant. The 3D-printed device generated as much or more thrust than existing droplet-emitting electrospray engines. With this technology, astronauts might quickly print an engine for a satellite without needing to wait for one to be sent up from Earth. 

“Using semiconductor manufacturing doesn’t match up with the idea of low-cost access to space. We want to democratize space hardware. In this work, we are proposing a way to make high-performance hardware with manufacturing techniques that are available to more players,” says Luis Fernando Velásquez-García, a principal research scientist in MIT’s Microsystems Technology Laboratories (MTL) and senior author of a paper describing the thrusters, which appears in Advanced Science.

He is joined on the paper by lead author Hyeonseok Kim, an MIT graduate student in mechanical engineering.

A modular approach

An electrospray engine has a reservoir of propellant that flows through microfluidic channels to a series of emitters. An electrostatic field is applied at the tip of each emitter, triggering an electrohydrodynamic effect that shapes the free surface of the liquid into a cone-shaped meniscus that ejects a stream of high-speed charged droplets from its apex, producing thrust.

The emitter tips need to be as sharp as possible to attain the electrohydrodynamic ejection of propellant at a low voltage. The device also requires a complex hydraulic system to store and regulate the flow of liquid, efficiently shuttling propellant through microfluidic channels.

The emitter array is composed of eight emitter modules. Each emitter module contains an array of four individual emitters that must work in unison, forming a larger system of interconnected modules.

“Using a one-size-fits-all fabrication approach doesn’t work because these subsystems are at different scales. Our key insight was to blend additive manufacturing methods to achieve the desired outcomes, then come up with a way to interface everything so the parts work together as efficiently as possible,” Velásquez-García says.

To accomplish this, the researchers utilized two different types of vat photo polymerization printing (VPP). VPP involves shining light onto a photosensitive resin, which solidifies to form 3D structures with smooth, high-resolution features.

The researchers fabricated the emitter modules using a VPP method called two-photon printing. This technique utilizes a highly focused laser beam to solidify resin in a precisely defined area, building a 3D structure one tiny brick, or voxel, at a time. This level of detail enabled them to produce extremely sharp emitter tips and narrow, uniform capillaries to carry propellant.

The emitter modules are fitted into a rectangular casing called a manifold block, which holds each in place and supplies the emitters with propellant. The manifold block also integrates the emitter modules with the extractor electrode that triggers propellant ejection from the emitter tips when a suitable voltage is applied. Fabricating the larger manifold block using two-photon printing would be infeasible because of the method’s low throughput and limited printing volume.

Instead, the researchers used a technique called digital light processing, which utilizes a chip-sized projector to shine light into the resin, solidifying one layer of the 3D structure at a time.

“Each technology works very well at a certain scale. Combining them, so they work together to produce one device, lets us take the best of each method,” Velásquez-García says.

Propelling performance

But 3D printing the electrospray engine components is only half the battle. The researchers also conducted chemical experiments to ensure the printing materials were compatible with the conductive liquid propellant. If not, the propellant might corrode the engine or cause it to crack, which is undesirable for hardware meant for long-term operation with little to no maintenance.

They also developed a method to clamp the separate parts together in a way that avoids misalignments which could hamper performance and ensures the device remains watertight.

In the end, their 3D-printed prototype was able to generate thrust more efficiently than larger, more expensive chemical rockets and outperformed existing droplet electrospray engines.

The researchers also investigated how adjusting the pressure of propellant and modulating the voltage applied to the engine affected the flow of droplets. Surprisingly, they achieved a wider range of thrust by modulating the voltage. This could eliminate the need for a complex network of pipes, valves, or pressure signals to regulate the flow of liquid, leading to a lighter, cheaper electrospray thruster that is also more efficient.

“We were able to show that a simpler thruster can achieve better results,” Velásquez-García says.

The researchers want to continue exploring the benefits of voltage modulation in future work. They also want to fabricate denser and larger arrays of emitter modules. In addition, they may explore the use of multiple electrodes to decouple the process of triggering of the electrohydrodynamic ejection of propellant from setting up the shape and speed of the emitted jet. In the long run, they also hope to demonstrate a CubeSat that utilizes a fully 3D-printed electrospray engine during its operation and deorbiting.

This research is funded, in part, by a MathWorks fellowship and the NewSat Project, and was carried out, in part, using MIT.nano facilities.

Imagine you’re a chef with a highly sought-after recipe. You write your top-secret instructions in a journal to ensure you remember them, but its location within the book is evident from the folds and tears on the edges of that often-referenced page.

Much like recipes in a cookbook, the instructions to execute programs are stored in specific locations within a computer’s physical memory. The standard security method — referred to as “address space layout randomization” (ASLR) — scatters this precious code to different places, but hackers can now find their new locations. Instead of hacking the software directly, they use approaches called microarchitectural side attacks that exploit hardware, identifying which memory areas are most frequently used. From there, they can use code to reveal passwords and make critical administrative changes in the system (also known as code-reuse attacks).

To enhance ASLR’s effectiveness, researchers from the MIT Computer Science and Artificial Intelligence Laboratory (CSAIL) have found a way to make these footprints vanish. Their “Oreo” method mitigates hardware attacks by removing randomized bits of addresses that lead to a program’s instructions before they’re translated to a physical location. It scrubs away traces of where code gadgets (or short sequences of instructions for specific tasks) are located before hackers can find them, efficiently enhancing security for operating systems like Linux.

Oreo has three layers, much like its tasty namesake. Between the virtual address space (which is used to reference program instructions) and the physical address space (where the code is located), Oreo adds a new “masked address space.” This re-maps code from randomized virtual addresses to fixed locations before it is executed within the hardware, making it difficult for hackers to trace the program’s original locations in the virtual address space through hardware attacks.

“We got the idea to structure it in three layers from Oreo cookies,” says Shixin Song, an MIT PhD student in electrical engineering and computer science (EECS) and CSAIL affiliate who is the lead author of a paper about the work. “Think of the white filling in the middle of that treat — our version of that is a layer that essentially whites out traces of gadget locations before they end up in the wrong hands.”

Senior author Mengjia Yan, an MIT associate professor of EECS and CSAIL principal investigator, believes Oreo’s masking abilities could make address space layout randomization more secure and reliable.

“ASLR was deployed in operating systems like Windows and Linux, but within the last decade, its security flaws have rendered it almost broken,” says Yan. “Our goal is to revive this mechanism in modern systems to defend microarchitecture attacks, so we’ve developed a software-hardware co-design mechanism that prevents leaking secret offsets that tell hackers where the gadgets are.”

The CSAIL researchers will present their findings about Oreo at the Network and Distributed System Security Symposium later this month.

Song and her coauthors evaluated how well Oreo could protect Linux by simulating hardware attacks in gem5, a platform commonly used to study computer architecture. The team found that it could prevent microarchitectural side attacks without hampering the software it protects.

Song observes that these experiments demonstrate how Oreo is a lightweight security upgrade for operating systems. “Our method introduces marginal hardware changes by only requiring a few extra storage units to store some metadata,” she says. “Luckily, it also has a minimal impact on software performance.”

While Oreo adds an extra step to program execution by scrubbing away revealing bits of data, it doesn’t slow down applications. This efficiency makes it a worthwhile security boost to ASLR for page-table-based virtual memory systems beyond Linux, such as those commonly found in major platforms such as Intel, AMD, and Arm.

In the future, the team will look to address speculative execution attacks — where hackers fool computers into predicting their next tasks, then steal the hidden data it leaves behind. Case in point: the infamous Meltdown/Spectre attacks in 2018.

To defend against speculative execution attacks, the team emphasizes that Oreo needs to be coupled with other security mechanisms (such as Spectre mitigations). This potential limitation extends to applying Oreo to larger systems.

“We think Oreo could be a useful software-hardware co-design platform for a broader type of applications,” says Yan. “In addition to targeting ASLR, we’re working on new methods that can help safeguard the critical crypto libraries widely used to safeguard information across people's network communication and cloud storage.”

Song and Yan wrote the paper with MIT EECS undergraduate researcher Joseph Zhang. The team’s work was supported, in part, by Amazon, the U.S. Air Force Office of Scientific Research, and ACE, a center within the ​​Semiconductor Research Corporation sponsored by the U.S. Defense Advanced Research Projects Agency (DARPA).

The ancient Greek philosopher and polymath Aristotle once concluded that the human heart is tri-chambered and that it was the single most important organ in the entire body, governing motion, sensation, and thought.

Today, we know that the human heart actually has four chambers and that the brain largely controls motion, sensation, and thought. But Aristotle was correct in observing that the heart is a vital organ, pumping blood to the rest of the body to reach other vital organs. When a life-threatening condition like heart failure strikes, the heart gradually loses the ability to supply other organs with enough blood and nutrients that enables them to function.

Researchers from MIT and Harvard Medical School recently published an open-access paper in Nature Communications Medicine, introducing a noninvasive deep learning approach that analyzes electrocardiogram (ECG) signals to accurately predict a patient’s risk of developing heart failure. In a clinical trial, the model showed results with accuracy comparable to gold-standard but more-invasive procedures, giving hope to those at risk of heart failure. The condition has recently seen a sharp increase in mortality, particularly among young adults, likely due to the growing prevalence of obesity and diabetes.

“This paper is a culmination of things I’ve talked about in other venues for several years,” says the paper’s senior author Collin Stultz, director of Harvard-MIT Program in Health Sciences and Technology and affiliate of the MIT Abdul Latif Jameel Clinic for Machine Learning in Health (Jameel Clinic). “The goal of this work is to identify those who are starting to get sick even before they have symptoms so that you can intervene early enough to prevent hospitalization.”

Of the heart’s four chambers, two are atria and two are ventricles — the right side of the heart has one atrium and one ventricle, and vice versa. In a healthy human heart, these chambers operate in a rhythmic synchrony: oxygen-poor blood flows into the heart via the right atrium. The right atrium contracts and the pressure generated pushes the blood into the right ventricle where the blood is then pumped into the lungs to be oxygenated. The oxygen-rich blood from the lungs then drains into the left atrium, which contracts, pumping the blood into the left ventricle. Another contraction follows, and the blood is ejected from the left ventricle via the aorta, flowing into veins branching out to the rest of the body.

“When the left atrial pressures become elevated, the blood drain from the lungs into the left atrium is impeded because it’s a higher-pressure system,” Stultz explains. In addition to being a professor of electrical engineering and computer science, Stultz is also a practicing cardiologist at Mass General Hospital (MGH). “The higher the pressure in the left atrium, the more pulmonary symptoms you develop — shortness of breath and so forth. Because the right side of the heart pumps blood through the pulmonary vasculature to the lungs, the elevated pressures in the left atrium translate to elevated pressures in the pulmonary vasculature.”

The current gold standard for measuring left atrial pressure is right heart catheterization (RHC), an invasive procedure that requires a thin tube (the catheter) attached to a pressure transmitter to be inserted into the right heart and pulmonary arteries. Physicians often prefer to assess risk noninvasively before resorting to RHC, by examining the patient’s weight, blood pressure, and heart rate.

But in Stultz’s view, these measures are coarse, as evidenced by the fact that one-in-four heart failure patients is readmitted to the hospital within 30 days. “What we are seeking is something that gives you information like that of an invasive device, other than a simple weight scale,” Stultz says.

In order to gather more comprehensive information on a patient’s heart condition, physicians typically use a 12-lead ECG, in which 10 adhesive patches are stuck onto the patient and linked with a machine that produces information from 12 different angles of the heart. However, 12-lead ECG machines are only accessible in clinical settings and they are also not typically used to assess heart failure risk.

Instead, what Stultz and other researchers propose is a Cardiac Hemodynamic AI monitoring System (CHAIS), a deep neural network capable of analyzing ECG data from a single lead — in other words, the patient only needs to have a single adhesive, commercially-available patch on their chest that they can wear outside of the hospital, untethered to a machine.

To compare CHAIS with the current gold standard, RHC, the researchers selected patients who were already scheduled for a catheterization and asked them to wear the patch 24 to 48 hours before the procedure, although patients were asked to remove the patch before catheterization took place. “When you get to within an hour-and-a-half [before the procedure], it’s 0.875, so it’s very, very good,” Stultz explains. “Thereby a measure from the device is equivalent and gives you the same information as if you were cathed in the next hour-and-a-half.”

“Every cardiologist understands the value of left atrial pressure measurements in characterizing cardiac function and optimizing treatment strategies for patients with heart failure,” says Aaron Aguirre SM '03, PhD '08, a cardiologist and critical care physician at MGH. “This work is important because it offers a noninvasive approach to estimating this essential clinical parameter using a widely available cardiac monitor.”

Aguirre, who completed a PhD in medical engineering and medical physics at MIT, expects that with further clinical validation, CHAIS will be useful in two key areas: first, it will aid in selecting patients who will most benefit from more invasive cardiac testing via RHC; and second, the technology could enable serial monitoring and tracking of left atrial pressure in patients with heart disease. “A noninvasive and quantitative method can help in optimizing treatment strategies in patients at home or in hospital,” Aguirre says. “I am excited to see where the MIT team takes this next.”

But the benefits aren’t just limited to patients — for patients with hard-to-manage heart failure, it becomes a challenge to keep them from being readmitted to the hospital without a permanent implant, taking up more space and more time of an already beleaguered and understaffed medical workforce.

The researchers have another ongoing clinical trial using CHAIS with MGH and Boston Medical Center that they hope to conclude soon to begin data analysis.

“In my view, the real promise of AI in health care is to provide equitable, state-of-the-art care to everyone, regardless of their socioeconomic status, background, and where they live,” Stultz says. “This work is one step towards realizing this goal.”

Should you grab your umbrella before you walk out the door? Checking the weather forecast beforehand will only be helpful if that forecast is accurate.

Spatial prediction problems, like weather forecasting or air pollution estimation, involve predicting the value of a variable in a new location based on known values at other locations. Scientists typically use tried-and-true validation methods to determine how much to trust these predictions.

But MIT researchers have shown that these popular validation methods can fail quite badly for spatial prediction tasks. This might lead someone to believe that a forecast is accurate or that a new prediction method is effective, when in reality that is not the case.

The researchers developed a technique to assess prediction-validation methods and used it to prove that two classical methods can be substantively wrong on spatial problems. They then determined why these methods can fail and created a new method designed to handle the types of data used for spatial predictions.

In experiments with real and simulated data, their new method provided more accurate validations than the two most common techniques. The researchers evaluated each method using realistic spatial problems, including predicting the wind speed at the Chicago O-Hare Airport and forecasting the air temperature at five U.S. metro locations.

Their validation method could be applied to a range of problems, from helping climate scientists predict sea surface temperatures to aiding epidemiologists in estimating the effects of air pollution on certain diseases.

“Hopefully, this will lead to more reliable evaluations when people are coming up with new predictive methods and a better understanding of how well methods are performing,” says Tamara Broderick, an associate professor in MIT’s Department of Electrical Engineering and Computer Science (EECS), a member of the Laboratory for Information and Decision Systems and the Institute for Data, Systems, and Society, and an affiliate of the Computer Science and Artificial Intelligence Laboratory (CSAIL).

Broderick is joined on the paper by lead author and MIT postdoc David R. Burt and EECS graduate student Yunyi Shen. The research will be presented at the International Conference on Artificial Intelligence and Statistics.

Evaluating validations

Broderick’s group has recently collaborated with oceanographers and atmospheric scientists to develop machine-learning prediction models that can be used for problems with a strong spatial component.

Through this work, they noticed that traditional validation methods can be inaccurate in spatial settings. These methods hold out a small amount of training data, called validation data, and use it to assess the accuracy of the predictor.

To find the root of the problem, they conducted a thorough analysis and determined that traditional methods make assumptions that are inappropriate for spatial data. Evaluation methods rely on assumptions about how validation data and the data one wants to predict, called test data, are related.

Traditional methods assume that validation data and test data are independent and identically distributed, which implies that the value of any data point does not depend on the other data points. But in a spatial application, this is often not the case.

For instance, a scientist may be using validation data from EPA air pollution sensors to test the accuracy of a method that predicts air pollution in conservation areas. However, the EPA sensors are not independent — they were sited based on the location of other sensors.

In addition, perhaps the validation data are from EPA sensors near cities while the conservation sites are in rural areas. Because these data are from different locations, they likely have different statistical properties, so they are not identically distributed.

“Our experiments showed that you get some really wrong answers in the spatial case when these assumptions made by the validation method break down,” Broderick says.

The researchers needed to come up with a new assumption.

Specifically spatial

Thinking specifically about a spatial context, where data are gathered from different locations, they designed a method that assumes validation data and test data vary smoothly in space.

For instance, air pollution levels are unlikely to change dramatically between two neighboring houses.

“This regularity assumption is appropriate for many spatial processes, and it allows us to create a way to evaluate spatial predictors in the spatial domain. To the best of our knowledge, no one has done a systematic theoretical evaluation of what went wrong to come up with a better approach,” says Broderick.

To use their evaluation technique, one would input their predictor, the locations they want to predict, and their validation data, then it automatically does the rest. In the end, it estimates how accurate the predictor’s forecast will be for the location in question. However, effectively assessing their validation technique proved to be a challenge.

“We are not evaluating a method, instead we are evaluating an evaluation. So, we had to step back, think carefully, and get creative about the appropriate experiments we could use,” Broderick explains.

First, they designed several tests using simulated data, which had unrealistic aspects but allowed them to carefully control key parameters. Then, they created more realistic, semi-simulated data by modifying real data. Finally, they used real data for several experiments.

Using three types of data from realistic problems, like predicting the price of a flat in England based on its location and forecasting wind speed, enabled them to conduct a comprehensive evaluation. In most experiments, their technique was more accurate than either traditional method they compared it to.

In the future, the researchers plan to apply these techniques to improve uncertainty quantification in spatial settings. They also want to find other areas where the regularity assumption could improve the performance of predictors, such as with time-series data.

This research is funded, in part, by the National Science Foundation and the Office of Naval Research.

The push to bring manufacturing back to the U.S. is running up against an unfortunate truth: The processes for making many critical materials today create toxic byproducts and other environmental hazards. That’s true for commonly used industrial metals like nickel and titanium, as well as specialty minerals, materials, and coatings that go into batteries, advanced electronics, and defense applications.

Now 6K, founded by former MIT research scientist Kamal Hadidi, is using a new production process to bring critical materials production back to America without the toxic byproducts.

The company is actively scaling its microwave plasma technology, which it calls UniMelt, to transform the way critical minerals are processed, creating new domestic supply chains in the process. UniMelt uses beams of tightly controlled thermal plasma to melt or vaporize precursor materials into particles with precise sizes and crystalline phases.

The technology converts metals, such as titanium, nickel, and refractory alloys, into particles optimized for additive manufacturing for a range of industrial applications. It is also being used to create battery materials for electric vehicles, grid infrastructure, and data centers.

“The markets and critical materials we are focused on are important for not just economic reasons but also U.S. national security, because the bulk of these materials are manufactured today in nonfriendly countries,” 6K CEO Saurabh Ullal says. “Now, the [U.S. government] and our growing customer base can leverage this technology invented at MIT to make the U.S. less dependent on these nonfriendly countries, ensuring supply chain independence now and in the future.”

Named after the 6,000-degree temperature of its plasma, 6K is currently selling its high-performance metal powders to parts manufacturers as well as defense, automotive, medical, and oil and gas companies for use in applications from engine components and medical implants to rockets. To scale its battery materials business, 6K is also building a 100,000-square-foot production facility in Jackson, Tennessee, which will begin construction later this year.

A weekend project

Between 1994 and 2007, Hadidi worked at the Plasma Science and Fusion Center (PFSC), where he developed plasma technologies for a range of applications, including hydrogen production, fuel reforming, and detecting environmental toxins. His first company was founded in 2000 out of the PFSC to detect mercury in coal-fired power plants’ smokestacks.

“I loved working at MIT,” Hadidi says. “It’s an amazing place that really challenges you. Just being there is so stimulating because everyone’s trying to come up with new solutions and connect dots between different fields.”

Hadidi also began using high-frequency microwave plasmas to create nanomaterials for use in optical applications. He wasn’t a materials expert, so he collaborated with Professor Eric Jordan, a materials synthesis expert from the University of Connecticut, and the researchers started working on nights and weekends in the PSFC to develop the idea further, eventually patenting the technology.

Hadidi officially founded the company as Amastan in 2007, exploring the use of his microwave plasma technology, later named UniMelt for “uniform melt state process,” to make a host of different materials as part of a government grant he and Jordan received.

The researchers soon realized the microwave plasma technology had several advantages over traditional production techniques for certain materials. For one, it could eliminate several high-energy steps of conventional processes, reducing production times from days to hours in some cases. For batteries and certain critical minerals, the process also works with recycled feedstocks. Amastan was renamed 6K in 2019.

Early on, Hadidi produced metal powders used in additive manufacturing through a process called spheroidization, which results in dense, spherical powders that flow well and make high-performance 3D-printed parts.

Following another grant, Hadidi explored methods for producing a type of battery cathode made from lithium, nickel, manganese, and cobalt (NMC). The standard process for making NMCs involved chemical synthesis, precipitation, heat treatment, and a lot of water. 6K is able to reduce many of those steps, speeding up production and lowering costs while also being more sustainable.

“Our technology completely eliminates toxic waste and recycles all of the byproducts back through the process to utilize everything, including water,” Ullal says.

Scaling domestic production

Today, 6K’s additive manufacturing arm operates out of a factory in Pennsylvania. The company’s critical minerals processing, refining, and recycling systems can produce about 400 tons of material per year and can be used to make more than a dozen types of metal powders. The company also has 33,000-square-foot battery center in North Andover, Massachusetts, where it produces battery cathode materials for its energy storage and mobility customers.

The Tennessee facility will be used to produce battery cathode materials and represents a massive step up in throughput. The company says it will be able to produce 13,000 tons of material annually when construction is complete next year.

“I’m happy if what I started brings something positive to society, and I’m extremely thankful to all the people that helped me,” says Hadidi, who left the company in 2019. “I’m an entrepreneur at heart. I like to make things. But that doesn’t mean I always succeed. It’s personally very satisfying to see this make an impact.”

The 6K team says its technology can also create a variety of specialty ceramics, advanced coatings, and nanoengineered materials. They say it may also be used to eliminate PFAS, or “forever chemicals,” though that work is at an early stage.

The company recently received a grant to demonstrate a process for recycling critical materials from military depots to produce aerospace and defense products, creating a new value stream for these materials that would otherwise deteriorate or go to landfill. That work is consistent with the company’s motto, “We take nothing from the ground and put nothing into the ground.”

The company’s additive division recently received a $23.4 Defense Production Act grant “that will enable us to double processing capacity in the next three years,” Ullal says. “The next step is to scale battery materials production to the tens of thousands of tons per year. At this point, it’s a scale-up of known processes, and we just need to execute. The idea of creating a circular economy is near and dear to us because that’s how we’ve built this company and that’s how we generate value: addressing our U.S. national security concerns and protecting the planet as well.”

A new technology developed at MIT enables scientists to label proteins across millions of individual cells in fully intact 3D tissues with unprecedented speed, uniformity, and versatility. Using the technology, the team was able to richly label large tissue samples in a single day. In their new study in Nature Biotechnology, they also demonstrate that the ability to label proteins with antibodies at the single-cell level across large tissue samples can reveal insights left hidden by other widely used labeling methods.

Profiling the proteins that cells are making is a staple of studies in biology, neuroscience, and related fields because the proteins a cell is expressing at a given moment can reflect the functions the cell is trying to perform or its response to its circumstances, such as disease or treatment. As much as microscopy and labeling technologies have advanced, enabling innumerable discoveries, scientists have still lacked a reliable and practical way of tracking protein expression at the level of millions of densely packed individual cells in whole, 3D intact tissues. Often confined to thin tissue sections under slides, scientists therefore haven’t had tools to thoroughly appreciate cellular protein expression in the whole, connected systems in which it occurs.

“Conventionally, investigating the molecules within cells requires dissociating tissue into single cells or slicing it into thin sections, as light and chemicals required for analysis cannot penetrate deep into tissues. Our lab developed technologies such as CLARITY and SHIELD, which enable investigation of whole organs by rendering them transparent, but we now needed a way to chemically label whole organs to gain useful scientific insights,” says study senior author Kwanghun Chung, associate professor in The Picower Institute for Learning and Memory, the departments of Chemical Engineering and Brain and Cognitive Sciences, and the Institute for Medical Engineering and Science at MIT. “If cells within a tissue are not uniformly processed, they cannot be quantitatively compared. In conventional protein labeling, it can take weeks for these molecules to diffuse into intact organs, making uniform chemical processing of organ-scale tissues virtually impossible and extremely slow.”

The new approach, called “CuRVE,” represents a major advance — years in the making — toward that goal by demonstrating a fundamentally new approach to uniformly processing large and dense tissues whole. In the study, the researchers explain how they overcame the technical barriers via an implementation of CuRVE called “eFLASH,” and provide copious vivid demonstrations of the technology, including how it yielded new neuroscience insights.

“This is a significant leap, especially in terms of the actual performance of the technology,” says co-lead author Dae Hee Yun PhD '24, a recent MIT graduate student who is now a senior application engineer at LifeCanvas Technologies, a startup company Chung founded to disseminate the tools his lab invents. The paper’s other lead author is Young-Gyun Park, a former MIT postdoc who’s now an assistant professor at KAIST in South Korea.

Clever chemistry

The fundamental reason why large, 3D tissue samples are hard to label uniformly is that antibodies seep into tissue very slowly, but are quick to bind to their target proteins. The practical effect of this speed mismatch is that simply soaking a brain in a bath of antibodies will mean that proteins are intensely well labeled on the outer edge of the tissue, but virtually none of the antibodies will find cells and proteins deeper inside.

To improve labeling, the team conceived of a way — the conceptual essence of CuRVE — to resolve the speed mismatch. The strategy was to continuously control the pace of antibody binding while at the same time speeding up antibody permeation throughout the tissue. To figure out how this could work and to optimize the approach, they built and ran a sophisticated computational simulation that enabled them to test different settings and parameters, including different binding rates and tissue densities and compositions.

Then they set out to implement their approach in real tissues. Their starting point was a previous technology, called “SWITCH,” in which Chung’s lab devised a way of temporarily turning off antibody binding, letting the antibodies permeate the tissue, and then turning binding back on. As well as it worked, Yun says, the team realized there could be substantial improvements if antibody binding speed could be controlled constantly, but the chemicals used in SWITCH were too harsh for such ongoing treatment. So the team screened a library of similar chemicals to find one that could more subtly and continuously throttle antibody binding speed. They found that deoxycholic acid was an ideal candidate. Using that chemical, the team could not only modulate antibody binding by varying the chemical’s concentration, but also by varying the labeling bath’s pH (or acidity).

Meanwhile, to speed up antibody movement through tissues, the team used another prior technology invented in the Chung Lab: stochastic electrotransport. That technology accelerates the dispersion of antibodies through tissue by applying electric fields.

Implementing this eFLASH system of accelerated dispersion with continuously modifiable binding speed produced the wide variety of labeling successes demonstrated in the paper. In all, the team reported using more than 60 different antibodies to label proteins in cells across large tissue samples.

Notably, each of these specimens was labeled within a day, an “ultra-fast” speed for whole, intact organs, the authors say. Moreover, different preparations did not require new optimization steps.

Valuable visualizations

Among the ways the team put eFLASH to the test was by comparing their labeling to another often-used method: genetically engineering cells to fluoresce when the gene for a protein of interest is being transcribed. The genetic method doesn’t require dispersing antibodies throughout tissue, but it can be prone to discrepancies because reporting gene transcription and actual protein production are not exactly the same thing. Yun added that while antibody labeling reliably and immediately reports on the presence of a target protein, the genetic method can be much less immediate and persistent, still fluorescing even when the actual protein is no longer present.

In the study the team employed both kinds of labeling simultaneously in samples. Visualizing the labels that way, they saw many examples in which antibody labeling and genetic labeling differed widely. In some areas of mouse brains, they found that two-thirds of the neurons expressing PV (a protein prominent in certain inhibitory neurons) according to antibody labeling, did not show any genetically-based fluorescence. In another example, only a tiny fraction of cells that reported expression via the genetic method of a protein called ChAT also reported it via antibody labeling. In other words, there were cases where genetic labeling both severely underreported or overreported protein expression compared to antibody labeling.

The researchers don’t mean to impugn the clear value of using the genetic reporting methods, but instead suggest that also using organ-wide antibody labeling, as eFLASH allows, can help put that data in a richer, more complete context. “Our discovery of large regionalized loss of PV-immunoreactive neurons in healthy adult mice and with high individual variability emphasizes the importance of holistic and unbiased phenotyping,” the authors write.

Or as Yun puts it, the two different kinds of labeling are “two different tools for the job.”

In addition to Yun, Park, and Chung, the paper’s other authors are Jae Hun Cho, Lee Kamentsky, Nicholas Evans, Nicholas DiNapoli, Katherine Xie, Seo Woo Choi, Alexandre Albanese, Yuxuan Tian, Chang Ho Sohn, Qiangge Zhang, Minyoung Kim, Justin Swaney, Webster Guan, Juhyuk Park, Gabi Drummond, Heejin Choi, Luzdary Ruelas, and Guoping Feng.

Funding for the study came from the Burroughs Wellcome Fund, the Searle Scholars Program, a Packard Award in Science and Engineering, a NARSAD Young Investigator Award, the McKnight Foundation, the Freedom Together Foundation, The Picower Institute for Learning and Memory, the NCSOFT Cultural Foundation, and the National Institutes of Health.

Sara Beery came to MIT as an assistant professor in MIT’s Department of Electrical Engineering and Computer Science (EECS) eager to focus on ecological challenges. She has fashioned her research career around the opportunity to apply her expertise in computer vision, machine learning, and data science to tackle real-world issues in conservation and sustainability. Beery was drawn to the Institute’s commitment to “computing for the planet,” and set out to bring her methods to global-scale environmental and biodiversity monitoring.

In the Pacific Northwest, salmon have a disproportionate impact on the health of their ecosystems, and their complex reproductive needs have attracted Beery’s attention. Each year, millions of salmon embark on a migration to spawn. Their journey begins in freshwater stream beds where the eggs hatch. Young salmon fry (newly hatched salmon) make their way to the ocean, where they spend several years maturing to adulthood. As adults, the salmon return to the streams where they were born in order to spawn, ensuring the continuation of their species by depositing their eggs in the gravel of the stream beds. Both male and female salmon die shortly after supplying the river habitat with the next generation of salmon. 

Throughout their migration, salmon support a wide range of organisms in the ecosystems they pass through. For example, salmon bring nutrients like carbon and nitrogen from the ocean upriver, enhancing their availability to those ecosystems. In addition, salmon are key to many predator-prey relationships: They serve as a food source for various predators, such as bears, wolves, and birds, while helping to control other populations, like insects, through predation. After they die from spawning, the decomposing salmon carcasses also replenish valuable nutrients to the surrounding ecosystem. The migration of salmon not only sustains their own species but plays a critical role in the overall health of the rivers and oceans they inhabit. 

At the same time, salmon populations play an important role both economically and culturally in the region. Commercial and recreational salmon fisheries contribute significantly to the local economy. And for many Indigenous peoples in the Pacific northwest, salmon hold notable cultural value, as they have been central to their diets, traditions, and ceremonies. 

Monitoring salmon migration

Increased human activity, including overfishing and hydropower development, together with habitat loss and climate change, have had a significant impact on salmon populations in the region. As a result, effective monitoring and management of salmon fisheries is important to ensure balance among competing ecological, cultural, and human interests. Accurately counting salmon during their seasonal migration to their natal river to spawn is essential in order to track threatened populations, assess the success of recovery strategies, guide fishing season regulations, and support the management of both commercial and recreational fisheries. Precise population data help decision-makers employ the best strategies to safeguard the health of the ecosystem while accommodating human needs. Monitoring salmon migration is a labor-intensive and inefficient undertaking.

Beery is currently leading a research project that aims to streamline salmon monitoring using cutting-edge computer vision methods. This project fits within Beery’s broader research interest, which focuses on the interdisciplinary space between artificial intelligence, the natural world, and sustainability. Its relevance to fisheries management made it a good fit for funding from MIT’s Abdul Latif Jameel Water and Food Systems Lab (J-WAFS). Beery’s 2023 J-WAFS seed grant was the first research funding she was awarded since joining the MIT faculty.  

Historically, monitoring efforts relied on humans to manually count salmon from riverbanks using eyesight. In the past few decades, underwater sonar systems have been implemented to aid in counting the salmon. These sonar systems are essentially underwater video cameras, but they differ in that they use acoustics instead of light sensors to capture the presence of a fish. Use of this method requires people to set up a tent alongside the river to count salmon based on the output of a sonar camera that is hooked up to a laptop. While this system is an improvement to the original method of monitoring salmon by eyesight, it still relies significantly on human effort and is an arduous and time-consuming process. 

Automating salmon monitoring is necessary for better management of salmon fisheries. “We need these technological tools,” says Beery. “We can’t keep up with the demand of monitoring and understanding and studying these really complex ecosystems that we work in without some form of automation.”

In order to automate counting of migrating salmon populations in the Pacific Northwest, the project team, including Justin Kay, a PhD student in EECS, has been collecting data in the form of videos from sonar cameras at different rivers. The team annotates a subset of the data to train the computer vision system to autonomously detect and count the fish as they migrate. Kay describes the process of how the model counts each migrating fish: “The computer vision algorithm is designed to locate a fish in the frame, draw a box around it, and then track it over time. If a fish is detected on one side of the screen and leaves on the other side of the screen, then we count it as moving upstream.” On rivers where the team has created training data for the system, it has produced strong results, with only 3 to 5 percent counting error. This is well below the target that the team and partnering stakeholders set of no more than a 10 percent counting error. 

Testing and deployment: Balancing human effort and use of automation

The researchers’ technology is being deployed to monitor the migration of salmon on the newly restored Klamath River. Four dams on the river were recently demolished, making it the largest dam removal project in U.S. history. The dams came down after a more than 20-year-long campaign to remove them, which was led by Klamath tribes, in collaboration with scientists, environmental organizations, and commercial fishermen. After the removal of the dams, 240 miles of the river now flow freely and nearly 800 square miles of habitat are accessible to salmon. Beery notes the almost immediate regeneration of salmon populations in the Klamath River: “I think it was within eight days of the dam coming down, they started seeing salmon actually migrate upriver beyond the dam.” In a collaboration with California Trout, the team is currently processing new data to adapt and create a customized model that can then be deployed to help count the newly migrating salmon.

One challenge with the system revolves around training the model to accurately count the fish in unfamiliar environments with variations such as riverbed features, water clarity, and lighting conditions. These factors can significantly alter how the fish appear on the output of a sonar camera and confuse the computer model. When deployed in new rivers where no data have been collected before, like the Klamath, the performance of the system degrades and the margin of error increases substantially to 15-20 percent. 

The researchers constructed an automatic adaptation algorithm within the system to overcome this challenge and create a scalable system that can be deployed to any site without human intervention. This self-initializing technology works to automatically calibrate to the new conditions and environment to accurately count the migrating fish. In testing, the automatic adaptation algorithm was able to reduce the counting error down to the 10 to 15 percent range. The improvement in counting error with the self-initializing function means that the technology is closer to being deployable to new locations without much additional human effort. 

Enabling real-time management with the “Fishbox”

Another challenge faced by the research team was the development of an efficient data infrastructure. In order to run the computer vision system, the video produced by sonar cameras must be delivered via the cloud or by manually mailing hard drives from a river site to the lab. These methods have notable drawbacks: a cloud-based approach is limited due to lack of internet connectivity in remote river site locations, and shipping the data introduces problems of delay. 

Instead of relying on these methods, the team has implemented a power-efficient computer, coined the “Fishbox,” that can be used in the field to perform the processing. The Fishbox consists of a small, lightweight computer with optimized software that fishery managers can plug into their existing laptops and sonar cameras. The system is then capable of running salmon counting models directly at the sonar sites without the need for internet connectivity. This allows managers to make hour-by-hour decisions, supporting more responsive, real-time management of salmon populations.

Community development

The team is also working to bring a community together around monitoring for salmon fisheries management in the Pacific Northwest. “It’s just pretty exciting to have stakeholders who are enthusiastic about getting access to [our technology] as we get it to work and having a tighter integration and collaboration with them,” says Beery. “I think particularly when you’re working on food and water systems, you need direct collaboration to help facilitate impact, because you're ensuring that what you develop is actually serving the needs of the people and organizations that you are helping to support.”

This past June, Beery’s lab organized a workshop in Seattle that convened nongovernmental organizations, tribes, and state and federal departments of fish and wildlife to discuss the use of automated sonar systems to monitor and manage salmon populations. Kay notes that the workshop was an “awesome opportunity to have everybody sharing different ways that they're using sonar and thinking about how the automated methods that we’re building could fit into that workflow.” The discussion continues now via a shared Slack channel created by the team, with over 50 participants. Convening this group is a significant achievement, as many of these organizations would not otherwise have had an opportunity to come together and collaborate. 

Looking forward

As the team continues to tune the computer vision system, refine their technology, and engage with diverse stakeholders — from Indigenous communities to fishery managers — the project is poised to make significant improvements to the efficiency and accuracy of salmon monitoring and management in the region. And as Beery advances the work of her MIT group, the J-WAFS seed grant is helping to keep challenges such as fisheries management in her sights.  

“The fact that the J-WAFS seed grant existed here at MIT enabled us to continue to work on this project when we moved here,” comments Beery, adding “it also expanded the scope of the project and allowed us to maintain active collaboration on what I think is a really important and impactful project.” 

As J-WAFS marks its 10th anniversary this year, the program aims to continue supporting and encouraging MIT faculty to pursue innovative projects that aim to advance knowledge and create practical solutions with real-world impacts on global water and food system challenges. 

Human activities continue to pump billions of tons of carbon dioxide into the atmosphere each year, raising global temperatures and driving extreme weather events. As countries grapple with climate impacts and ways to significantly reduce carbon emissions, there have been various efforts to advance carbon dioxide removal (CDR) technologies that directly remove carbon dioxide from the air and sequester it for long periods of time.

Unlike carbon capture and storage technologies, which are designed to remove carbon dioxide at point sources such as fossil-fuel plants, CDR aims to remove carbon dioxide molecules that are already circulating in the atmosphere.

A new report by the American Physical Society and led by an MIT physicist provides an overview of the major experimental CDR approaches and determines their fundamental physical limits. The report focuses on methods that have the biggest potential for removing carbon dioxide, at the scale of gigatons per year, which is the magnitude that would be required to have a climate-stabilizing impact.

The new report was commissioned by the American Physical Society's Panel on Public Affairs, and appeared last week in the journal PRX. The report was chaired by MIT professor of physics Washington Taylor, who spoke with MIT News about CDR’s physical limitations and why it’s worth pursuing in tandem with global efforts to reduce carbon emissions.

Q: What motivated you to look at carbon dioxide removal systems from a physical science perspective?

A: The number one thing driving climate change is the fact that we’re taking carbon that has been stuck in the ground for 100 million years, and putting it in the atmosphere, and that’s causing warming. In the last few years there’s been a lot of interest both by the government and private entities in finding technologies to directly remove the CO2 from the air.

How to manage atmospheric carbon is the critical question in dealing with our impact on Earth’s climate. So, it’s very important for us to understand whether we can affect the carbon levels not just by changing our emissions profile but also by directly taking carbon out of the atmosphere. Physics has a lot to say about this because the possibilities are very strongly constrained by thermodynamics, mass issues, and things like that.

Q: What carbon dioxide removal methods did you evaluate?

A: They’re all at an early stage. It's kind of the Wild West out there in terms of the different ways in which companies are proposing to remove carbon from the atmosphere. In this report, we break down CDR processes into two classes: cyclic and once-through.

Imagine we are in a boat that has a hole in the hull and is rapidly taking on water. Of course, we want to plug the hole as quickly as we can. But even once we have fixed the hole, we need to get the water out so we aren't in danger of sinking or getting swamped. And this is particularly urgent if we haven't completely fixed the hole so we still have a slow leak. Now, imagine we have a couple of options for how to get the water out so we don’t sink.

The first is a sponge that we can use to absorb water, that we can then squeeze out and reuse. That’s a cyclic process in the sense that we have some material that we’re using over and over. There are cyclic CDR processes like chemical “direct air capture” (DAC), which acts basically like a sponge. You set up a big system with fans that blow air past some material that captures carbon dioxide. When the material is saturated, you close off the system and then use energy to essentially squeeze out the carbon and store it in a deep repository. Then you can reuse the material, in a cyclic process.

The second class of approaches is what we call “once-through.” In the boat analogy, it would be as if you try to fix the leak using cartons of paper towels. You let them saturate and then throw them overboard, and you use each roll once.

There are once-through CDR approaches, like enhanced rock weathering, that are designed to accelerate a natural process, by which certain rocks, when exposed to air, will absorb carbon from the atmosphere. Worldwide, this natural rock weathering is estimated to remove about 1 gigaton of carbon each year. “Enhanced rock weathering” is a CDR approach where you would dig up a lot of this rock, grind it up really small, to less than the width of a human hair, to get the process to happen much faster. The idea is, you dig up something, spread it out, and absorb CO2 in one go.

The key difference between these two processes is that the cyclic process is subject to the second law of thermodynamics and there’s an energy constraint. You can set an actual limit from physics, saying any cyclic process is going to take a certain amount of energy, and that cannot be avoided. For example, we find that for cyclic direct-air-capture (DAC) plants, based on second law limits, the absolute minimum amount of energy you would need to capture a gigaton of carbon is comparable to the total yearly electric energy consumption of the state of Virginia. Systems currently under development use at least three to 10 times this much energy on a per ton basis (and capture tens of thousands, not billions, of tons). Such systems also need to move a lot of air; the air that would need to pass through a DAC system to capture a gigaton of CO2 is comparable to the amount of air that passes through all the air cooling systems on the planet.

On the other hand, if you have a once-through process, you could in some respects avoid the energy constraint, but now you’ve got a materials constraint due to the central laws of chemistry. For once-through processes like enhanced rock weathering, that means that if you want to capture a gigaton of CO2, roughly speaking, you’re going to need a billion tons of rock.

So, to capture gigatons of carbon through engineered methods requires tremendous amounts of physical material, air movement, and energy. On the other hand, everything we’re doing to put that CO2 in the atmosphere is extensive too, so large-scale emissions reductions face comparable challenges.

Q: What does the report conclude, in terms of whether and how to remove carbon dioxide from the atmosphere?

A: Our initial prejudice was, CDR is just going to take so much energy, and there’s no way around that because of the second law of thermodynamics, regardless of the method.

But as we discussed, there is this nuance about cyclic versus once-through systems. And there are two points of view that we ended up threading a needle between. One is the view that CDR is a silver bullet, and we’ll just do CDR and not worry about emissions — we’ll just suck it all out of the atmosphere. And that’s not the case. It will be really expensive, and will take a lot of energy and materials to do large-scale CDR. But there’s another view, where people say, don’t even think about CDR. Even thinking about CDR will compromise our efforts toward emissions reductions. The report comes down somewhere in the middle, saying that CDR is not a magic bullet, but also not a no-go.

If we are serious about managing climate change, we will likely want substantial CDR in addition to aggressive emissions reductions. The report concludes that research and development on CDR methods should be selectively and prudently pursued despite the expected cost and energy and material requirements.

At a policy level, the main message is that we need an economic and policy framework that incentivizes emissions reductions and CDR in a common framework; this would naturally allow the market to optimize climate solutions. Since in many cases it is much easier and cheaper to cut emissions than it will likely ever be to remove atmospheric carbon, clearly understanding the challenges of CDR should help motivate rapid emissions reductions.

For me, I’m optimistic in the sense that scientifically we understand what it will take to reduce emissions and to use CDR to bring CO2 levels down to a slightly lower level. Now, it’s really a societal and economic problem. I think humanity has the potential to solve these problems. I hope that we can find common ground so that we can take actions as a society that will benefit both humanity and the broader ecosystems on the planet, before we end up having bigger problems than we already have. 

In India, many kids who work in retail markets have good math skills: They can quickly perform a range of calculations to complete transactions. But as a new study shows, these kids often perform much worse on the same kinds of problems as they are taught in the classroom. This happens even though many of these students still attend school or attended school through 7th or 8th grades.

Conversely, the study also finds, Indian students who are still enrolled in school and don’t have jobs do better on school-type math problems, but they often fare poorly at the kinds of problems that occur in marketplaces.

Overall, both the “market kids” and the “school kids” struggle with the approach the other group is proficient in, raising questions about how to help both groups learn math more comprehensively.

“For the school kids, they do worse when you go from an abstract problem to a concrete problem,” says MIT economist Esther Duflo, co-author of a new paper detailing the study’s results. “For the market kids, it’s the opposite.”

Indeed, the kids with jobs who are also in school “underperform despite being extraordinarily good at mental math,” says Abhijit Banerjee an MIT economist and another co-author of the paper. “That for me was always the revelation, that the one doesn’t translate into the other.”

The paper, “Children’s arithmetic skills do not transfer between applied and academic math,” is published today in Nature. The authors are Banerjee, the Ford Professor of Economics at MIT; Swati Bhattacharjee of the newspaper Ananda Bazar Patrika, in Kolkata, India; Raghabendra Chattopadhyay of the Indian Institute of Management in Kolkata; Duflo, the Abdul Latif Jameel Professor of Poverty Alleviation and Development Economics at MIT; Alejandro J. Ganimian, a professor of applied psychology and economics at New York University; Kailash Rajaha, a doctoral candidate in economics at MIT; and Elizabeth S. Spelke, a professor of psychology at Harvard University.

Duflo and Banerjee shared the Nobel Prize in Economics in 2019 and are co-founders of MIT’s Jameel Abdul Latif Poverty Action Lab (J-PAL), a global leader in development economics.

Three experiments

The study consists largely of three data-collection exercises with some embedded experiments. The first one shows that 201 kids working in markets in Kolkata do have good math skills. For instance, a researcher, posing as an ordinary shopper, would ask for the cost of 800 grams of potatoes sold at 20 rupees per kilogram, then ask for the cost of 1.4 kilograms of onions sold at 15 rupees per kilo. They would request the combined answer — 37 rupees —  then hand the market worker a 200 rupee note and collect 163 rupees back. All told, the kids working in markets correctly solved this kind of problem from 95 to 98 percent of the time by the second try.

However, when the working children were pulled aside (with their parents’ permission) and given a standardized Indian national math test, just 32 percent could correctly divide a three-digit number by a one-digit number, and just 54 percent could correctly subtract a two-digit number from another two-digit number two times. Clearly, the kids’ skills were not yielding classroom results.

The researchers then conducted a second study with 400 kids working in markets in Delhi, which replicated the results: Working kids had a strong ability to handle market transactions, but only about 15 percent of the ones also in school were at average proficiency in math.

In the second study, the researchers also asked the reverse question: How do students doing well in school fare at market math problems? Here, with 200 students from 17 Delhi schools who do not work in markets, they found that 96 percent of the students could solve typical problems with a pencil, paper, unlimited time, and one opportunity to self-correct. But when the students had to solve the problems in a make-believe “market” setting, that figure dropped to just 60 percent. The students had unlimited time and access to paper and pencil, so that figure may actually overestimate how they would fare in a market.

Finally, in a third study, conducted in Delhi with over 200 kids, the researchers compared the performances of both “market” and “school” kids again on numerous math problems in varying conditions. While 85 percent of the working kids got the right answer to a market transaction problem, only 10 percent of nonworking kids correctly answered a question of similar difficulty, when faced with limited time and with no aids like pencil and paper. However, given the same division and subtraction problems, but with pencil and paper, 59 percent of nonmarket kids got them right, compared to 45 percent of market kids.

To further evaluate market kids and school kids on a level playing field, the researchers then presented each group with a word problem about a boy going to the market and buying two vegetables. Roughly one-third of the market kids were able to solve this without any aid, while fewer than 1 percent of the school kids did.

Why might the performance of the nonworking students decline when given a problem in market conditions?

“They learned an algorithm but didn’t understand it,” Banerjee says.

Meanwhile, the market kids seemed to use certain tactics to handle retail transactions. For one thing, they appear to use rounding well. Take a problem like 43 times 11. To handle that intuitively, you might multiply 43 times 10, and then add 43, for the final answer of 473. This appears to be what they are doing.

“The market kids are able to exploit base 10, so they do better on base 10 problems,” Duflo says. “The school kids have no idea. It makes no difference to them. The market kids may have additional tricks of this sort that we did not see.” On the other hand, the school kids had a better grasp of formal written methods of division, subtraction, and more.

Going farther in school

The findings raise a significant point about students skills and academic progress. While it is a good thing that the kids with market jobs are proficient at generating rapid answers, it would likely be better for the long-term futures if they also did well in school and wound up with a high school degree or better. Finding a way to cross the divide between informal and formal ways of tackling math problems, then, could notably help some Indian children.

The fact that such a divide exists, meanwhile, suggests some new approaches could be tried in the classroom.

Banerjee, for one, suspects that part of the issue is a classroom process making it seem as if there is only one true route to funding an arithmetic answer. Instead, he believes, following the work of co-author Spelke, that helping students reason their way to an approximation of the right answer can help them truly get a handle on what is needed to solve these types of problems.

Even so, Duflo adds, “We don’t want to blame the teachers. It’s not their fault. They are given a strict curriculum to follow, and strict methods to follow.”

That still leaves open the question of what to change, in concrete classroom terms. That topic, it happens, is something the research group is in the process of weighing, as they consider new experiments that might address it directly. The current finding, however, makes clear progress would be useful.

“These findings highlight the importance of educational curricula that bridge the gap between intuitive and formal mathematics,” the authors state in the paper.

Support for the research was provided, in part, by the Abdul Latif Jameel Poverty Action Lab’s Post-Primary Education Initiative, the Foundation Blaise Pascal, and the AXA Research Fund. 

Superconducting materials are similar to the carpool lane in a congested interstate. Like commuters who ride together, electrons that pair up can bypass the regular traffic, moving through the material with zero friction.

But just as with carpools, how easily electron pairs can flow depends on a number of conditions, including the density of pairs that are moving through the material. This “superfluid stiffness,” or the ease with which a current of electron pairs can flow, is a key measure of a material’s superconductivity.

Physicists at MIT and Harvard University have now directly measured superfluid stiffness for the first time in “magic-angle” graphene — materials that are made from two or more atomically thin sheets of graphene twisted with respect to each other at just the right angle to enable a host of exceptional properties, including unconventional superconductivity.

This superconductivity makes magic-angle graphene a promising building block for future quantum-computing devices, but exactly how the material superconducts is not well-understood. Knowing the material’s superfluid stiffness will help scientists identify the mechanism of superconductivity in magic-angle graphene.

The team’s measurements suggest that magic-angle graphene’s superconductivity is primarily governed by quantum geometry, which refers to the conceptual “shape” of quantum states that can exist in a given material.

The results, which are reported today in the journal Nature, represent the first time scientists have directly measured superfluid stiffness in a two-dimensional material. To do so, the team developed a new experimental method which can now be used to make similar measurements of other two-dimensional superconducting materials.

“There’s a whole family of 2D superconductors that is waiting to be probed, and we are really just scratching the surface,” says study co-lead author Joel Wang, a research scientist in MIT’s Research Laboratory of Electronics (RLE).

The study’s co-authors from MIT’s main campus and MIT Lincoln Laboratory include co-lead author and former RLE postdoc Miuko Tanaka as well as Thao Dinh, Daniel Rodan-Legrain, Sameia Zaman, Max Hays, Bharath Kannan, Aziza Almanakly, David Kim, Bethany Niedzielski, Kyle Serniak, Mollie Schwartz, Jeffrey Grover, Terry Orlando, Simon Gustavsson, Pablo Jarillo-Herrero, and William D. Oliver, along with Kenji Watanabe and Takashi Taniguchi of the National Institute for Materials Science in Japan.

Magic resonance

Since its first isolation and characterization in 2004, graphene has proven to be a wonder substance of sorts. The material is effectively a single, atom-thin sheet of graphite consisting of a precise, chicken-wire lattice of carbon atoms. This simple configuration can exhibit a host of superlative qualities in terms of graphene’s strength, durability, and ability to conduct electricity and heat.

In 2018, Jarillo-Herrero and colleagues discovered that when two graphene sheets are stacked on top of each other, at a precise “magic” angle, the twisted structure — now known as magic-angle twisted bilayer graphene, or MATBG — exhibits entirely new properties, including superconductivity, in which electrons pair up, rather than repelling each other as they do in everyday materials. These so-called Cooper pairs can form a superfluid, with the potential to superconduct, meaning they could move through a material as an effortless, friction-free current.

“But even though Cooper pairs have no resistance, you have to apply some push, in the form of an electric field, to get the current to move,” Wang explains. “Superfluid stiffness refers to how easy it is to get these particles to move, in order to drive superconductivity.”

Today, scientists can measure superfluid stiffness in superconducting materials through methods that generally involve placing a material in a microwave resonator — a device which has a characteristic resonance frequency at which an electrical signal will oscillate, at microwave frequencies, much like a vibrating violin string. If a superconducting material is placed within a microwave resonator, it can change the device’s resonance frequency, and in particular, its “kinetic inductance,” by an amount that scientists can directly relate to the material’s superfluid stiffness.

However, to date, such approaches have only been compatible with large, thick material samples. The MIT team realized that to measure superfluid stiffness in atomically thin materials like MATBG would require a new approach.

“Compared to MATBG, the typical superconductor that is probed using resonators is 10 to 100 times thicker and larger in area,” Wang says. “We weren’t sure if such a tiny material would generate any measurable inductance at all.”

A captured signal

The challenge to measuring superfluid stiffness in MATBG has to do with attaching the supremely delicate material to the surface of the microwave resonator as seamlessly as possible.

“To make this work, you want to make an ideally lossless — i.e., superconducting — contact between the two materials,” Wang explains. “Otherwise, the microwave signal you send in will be degraded or even just bounce back instead of going into your target material.”

Will Oliver’s group at MIT has been developing techniques to precisely connect extremely delicate, two-dimensional materials, with the goal of building new types of quantum bits for future quantum-computing devices. For their new study, Tanaka, Wang, and their colleagues applied these techniques to seamlessly connect a tiny sample of MATBG to the end of an aluminum microwave resonator. To do so, the group first used conventional methods to assemble MATBG, then sandwiched the structure between two insulating layers of hexagonal boron nitride, to help maintain MATBG’s atomic structure and properties.

“Aluminum is a material we use regularly in our superconducting quantum computing research, for example, aluminum resonators to read out aluminum quantum bits (qubits),” Oliver explains. “So, we thought, why not make most of the resonator from aluminum, which is relatively straightforward for us, and then add a little MATBG to the end of it? It turned out to be a good idea.”

“To contact the MATBG, we etch it very sharply, like cutting through layers of a cake with a very sharp knife,” Wang says. “We expose a side of the freshly-cut MATBG, onto which we then deposit aluminum — the same material as the resonator — to make a good contact and form an aluminum lead.”

The researchers then connected the aluminum leads of the MATBG structure to the larger aluminum microwave resonator. They sent a microwave signal through the resonator and measured the resulting shift in its resonance frequency, from which they could infer the kinetic inductance of the MATBG.

When they converted the measured inductance to a value of superfluid stiffness, however, the researchers found that it was much larger than what conventional theories of superconductivity would have predicted. They had a hunch that the surplus had to do with MATBG’s quantum geometry — the way the quantum states of electrons correlate to one another.

“We saw a tenfold increase in superfluid stiffness compared to conventional expectations, with a temperature dependence consistent with what the theory of quantum geometry predicts,” Tanaka says. “This was a ‘smoking gun’ that pointed to the role of quantum geometry in governing superfluid stiffness in this two-dimensional material.”

“This work represents a great example of how one can use sophisticated quantum technology currently used in quantum circuits to investigate condensed matter systems consisting of strongly interacting particles,” adds Jarillo-Herrero.

This research was funded, in part, by the U.S. Army Research Office, the National Science Foundation, the U.S. Air Force Office of Scientific Research, and the U.S. Under Secretary of Defense for Research and Engineering. The work was carried out, in part, through the use of MIT.nano’s facilities.

A complementary study on magic-angle twisted trilayer graphene (MATTG), conducted by a collaboration between Philip Kim’s group at Harvard University and Jarillo-Herrero’s group at MIT appears in the same issue of Nature.

When people think about fiber optic cables, its usually about how they’re used for telecommunications and accessing the internet. But fiber optic cables — strands of glass or plastic that allow for the transmission of light — can be used for another purpose: imaging the ground beneath our feet.

MIT Department of Earth, Atmospheric and Planetary Sciences (EAPS) PhD student Hilary Chang recently used the MIT fiber optic cable network to successfully image the ground underneath campus using a method known as distributed acoustic sensing (DAS). By using existing infrastructure, DAS can be an efficient and effective way to understand ground composition, a critical component for assessing the seismic hazard of areas, or how at risk they are from earthquake damage.

“We were able to extract very nice, coherent waves from the surroundings, and then use that to get some information about the subsurface,” says Chang, the lead author of a recent paper describing her work that was co-authored with EAPS Principal Research Scientist Nori Nakata. 

Dark fibers

The MIT campus fiber optic system, installed from 2000 to 2003, services internal data transport between labs and buildings as well as external transport, such as the campus internet (MITNet). There are three major cable hubs on campus from which lines branch out into buildings and underground, much like a spiderweb.

The network allocates a certain number of strands per building, some of which are “dark fibers,” or cables that are not actively transporting information. Each campus fiber hub has redundant backbone cables between them so that, in the event of a failure, network transmission can switch to the dark fibers without loss of network services.

DAS can use existing telecommunication cables and ambient wavefields to extract information about the materials they pass through, making it a valuable tool for places like cities or the ocean floor, where conventional sensors can’t be deployed. Chang, who studies earthquake waveforms and the information we can extract from them, decided to try it out on the MIT campus.

In order to get access to the fiber optic network for the experiment, Chang reached out to John Morgante, a manager of infrastructure project engineering with MIT Information Systems and Technology (IS&T). Morgante has been at MIT since 1998 and was involved with the original project installing the fiber optic network, and was thus able to provide personal insight into selecting a route.

“It was interesting to listen to what they were trying to accomplish with the testing,” says Morgante. While IS&T has worked with students before on various projects involving the school’s network, he said that “in the physical plant area, this is the first that I can remember that we’ve actually collaborated on an experiment together.”

They decided on a path starting from a hub in Building 24, because it was the longest running path that was entirely underground; above-ground wires that cut through buildings wouldn’t work because they weren’t grounded, and thus were useless for the experiment. The path ran from east to west, beginning in Building 24, traveling under a section of Massachusetts Ave., along parts of Amherst and Vassar streets, and ending at Building W92.

“[Morgante] was really helpful,” says Chang, describing it as “a very good experience working with the campus IT team.”

Locating the cables

After renting an interrogator, a device that sends laser pulses to sense ambient vibrations along the fiber optic cables, Chang and a group of volunteers were given special access to connect it to the hub in Building 24. They let it run for five days.

To validate the route and make sure that the interrogator was working, Chang conducted a tap test, in which she hit the ground with a hammer several times to record the precise GPS coordinates of the cable. Conveniently, the underground route is marked by maintenance hole covers that serve as good locations to do the test. And, because she needed the environment to be as quiet as possible to collect clean data, she had to do it around 2 a.m.

“I was hitting it next to a dorm and someone yelled ‘shut up,’ probably because the hammer blows woke them up,” Chang recalls. “I was sorry.” Thankfully, she only had to tap at a few spots and could interpolate the locations for the rest.

During the day, Chang and her fellow students — Denzel Segbefia, Congcong Yuan, and Jared Bryan — performed an additional test with geophones, another instrument that detects seismic waves, out on Brigg’s Field where the cable passed under it to compare the signals. It was an enjoyable experience for Chang; when the data were collected in 2022, the campus was coming out of pandemic measures, with remote classes sometimes still in place. “It was very nice to have everyone on the field and do something with their hands,” she says.

The noise around us

Once Chang collected the data, she was able to see plenty of environmental activity in the waveforms, including the passing of cars, bikes, and even when the train that runs along the northern edge of campus made its nightly passes.

After identifying the noise sources, Chang and Nakata extracted coherent surface waves from the ambient noises and used the wave speeds associated with different frequencies to understand the properties of the ground the cables passed through. Stiffer materials allow fast velocities, while softer material slows it.

“We found out that the MIT campus is built on soft materials overlaying a relatively hard bedrock,” Chang says, which confirms previously known, albeit lower-resolution, information about the geology of the area that had been collected using seismometers.

Information like this is critical for regions that are susceptible to destructive earthquakes and other seismic hazards, including the Commonwealth of Massachusetts, which has experienced earthquakes as recently as this past week. Areas of Boston and Cambridge characterized by artificial fill during rapid urbanization are especially at risk due to its subsurface structure being more likely to amplify seismic frequencies and damage buildings. This non-intrusive method for site characterization can help ensure that buildings meet code for the correct seismic hazard level.

“Destructive seismic events do happen, and we need to be prepared,” she says.

From crafting complex code to revolutionizing the hiring process, generative artificial intelligence is reshaping industries faster than ever before — pushing the boundaries of creativity, productivity, and collaboration across countless domains.

Enter the MIT Generative AI Impact Consortium, a collaboration between industry leaders and MIT’s top minds. As MIT President Sally Kornbluth highlighted last year, the Institute is poised to address the societal impacts of generative AI through bold collaborations. Building on this momentum and established through MIT’s Generative AI Week and impact papers, the consortium aims to harness AI’s transformative power for societal good, tackling challenges before they shape the future in unintended ways.

“Generative AI and large language models [LLMs] are reshaping everything, with applications stretching across diverse sectors,” says Anantha Chandrakasan, dean of the School of Engineering and MIT’s chief innovation and strategy officer, who leads the consortium. “As we push forward with newer and more efficient models, MIT is committed to guiding their development and impact on the world.”

Chandrakasan adds that the consortium’s vision is rooted in MIT’s core mission. “I am thrilled and honored to help advance one of President Kornbluth’s strategic priorities around artificial intelligence,” he says. “This initiative is uniquely MIT — it thrives on breaking down barriers, bringing together disciplines, and partnering with industry to create real, lasting impact. The collaborations ahead are something we’re truly excited about.”

Developing the blueprint for generative AI’s next leap

The consortium is guided by three pivotal questions, framed by Daniel Huttenlocher, dean of the MIT Schwarzman College of Computing and co-chair of the GenAI Dean’s oversight group, that go beyond AI’s technical capabilities and into its potential to transform industries and lives:

1. How can AI-human collaboration create outcomes that neither could achieve alone?
2. What is the dynamic between AI systems and human behavior, and how do we maximize the benefits while steering clear of risks?
3. How can interdisciplinary research guide the development of better, safer AI technologies that improve human life?

Generative AI continues to advance at lightning speed, but its future depends on building a solid foundation. “Everybody recognizes that large language models will transform entire industries, but there's no strong foundation yet around design principles,” says Tim Kraska, associate professor of electrical engineering and computer science in the MIT Computer Science and Artificial Intelligence Laboratory (CSAIL) and co-faculty director of the consortium.

“Now is a perfect time to look at the fundamentals — the building blocks that will make generative AI more effective and safer to use,” adds Kraska.

"What excites me is that this consortium isn’t just academic research for the distant future — we’re working on problems where our timelines align with industry needs, driving meaningful progress in real time," says Vivek F. Farias, the Patrick J. McGovern (1959) Professor at the MIT Sloan School of Management, and co-faculty director of the consortium.

A “perfect match” of academia and industry

At the heart of the Generative AI Impact Consortium are six founding members: Analog Devices, The Coca-Cola Co., OpenAI, Tata Group, SK Telecom, and TWG Global. Together, they will work hand-in-hand with MIT researchers to accelerate breakthroughs and address industry-shaping problems.

The consortium taps into MIT’s expertise, working across schools and disciplines — led by MIT’s Office of Innovation and Strategy, in collaboration with the MIT Schwarzman College of Computing and all five of MIT’s schools.

“This initiative is the ideal bridge between academia and industry,” says Chandrakasan. “With companies spanning diverse sectors, the consortium brings together real-world challenges, data, and expertise. MIT researchers will dive into these problems to develop cutting-edge models and applications into these different domains.”

Industry partners: Collaborating on AI’s evolution

At the core of the consortium’s mission is collaboration — bringing MIT researchers and industry partners together to unlock generative AI’s potential while ensuring its benefits are felt across society.

Among the founding members is OpenAI, the creator of the generative AI chatbot ChatGPT.

“This type of collaboration between academics, practitioners, and labs is key to ensuring that generative AI evolves in ways that meaningfully benefit society,” says Anna Makanju, vice president of global impact at OpenAI, adding that OpenAI “is eager to work alongside MIT’s Generative AI Consortium to bridge the gap between cutting-edge AI research and the real-world expertise of diverse industries.”

The Coca-Cola Co. recognizes an opportunity to leverage AI innovation on a global scale. “We see a tremendous opportunity to innovate at the speed of AI and, leveraging The Coca-Cola Company's global footprint, make these cutting-edge solutions accessible to everyone,” says Pratik Thakar, global vice president and head of generative AI. “Both MIT and The Coca-Cola Company are deeply committed to innovation, while also placing equal emphasis on the legally and ethically responsible development and use of technology.”

For TWG Global, the consortium offers the ideal environment to share knowledge and drive advancements. “The strength of the consortium is its unique combination of industry leaders and academia, which fosters the exchange of valuable lessons, technological advancements, and access to pioneering research,” says Drew Cukor, head of data and artificial intelligence transformation. Cukor adds that TWG Global “is keen to share its insights and actively engage with leading executives and academics to gain a broader perspective of how others are configuring and adopting AI, which is why we believe in the work of the consortium.”

The Tata Group views the collaboration as a platform to address some of AI’s most pressing challenges. “The consortium enables Tata to collaborate, share knowledge, and collectively shape the future of generative AI, particularly in addressing urgent challenges such as ethical considerations, data privacy, and algorithmic biases,” says Aparna Ganesh, vice president of Tata Sons Ltd.

Similarly, SK Telecom sees its involvement as a launchpad for growth and innovation. Suk-geun (SG) Chung, SK Telecom executive vice president and chief AI global officer, explains, “Joining the consortium presents a significant opportunity for SK Telecom to enhance its AI competitiveness in core business areas, including AI agents, AI semiconductors, data centers (AIDC), and physical AI,” says Chung. “By collaborating with MIT and leveraging the SK AI R&D Center as a technology control tower, we aim to forecast next-generation generative AI technology trends, propose innovative business models, and drive commercialization through academic-industrial collaboration.”

Alan Lee, chief technology officer of Analog Devices (ADI), highlights how the consortium bridges key knowledge gaps for both his company and the industry at large. “ADI can’t hire a world-leading expert in every single corner case, but the consortium will enable us to access top MIT researchers and get them involved in addressing problems we care about, as we also work together with others in the industry towards common goals,” he says.

The consortium will host interactive workshops and discussions to identify and prioritize challenges. “It’s going to be a two-way conversation, with the faculty coming together with industry partners, but also industry partners talking with each other,” says Georgia Perakis, the John C Head III Dean (Interim) of the MIT Sloan School of Management and professor of operations management, operations research and statistics, who serves alongside Huttenlocher as co-chair of the GenAI Dean’s oversight group.

Preparing for the AI-enabled workforce of the future

With AI poised to disrupt industries and create new opportunities, one of the consortium’s core goals is to guide that change in a way that benefits both businesses and society.

“When the first commercial digital computers were introduced [the UNIVAC was delivered to the U.S. Census Bureau in 1951], people were worried about losing their jobs,” says Kraska. “And yes, jobs like large-scale, manual data entry clerks and human ‘computers,’ people tasked with doing manual calculations, largely disappeared over time. But the people impacted by those first computers were trained to do other jobs.”

The consortium aims to play a key role in preparing the workforce of tomorrow by educating global business leaders and employees on generative AI evolving uses and applications. With the pace of innovation accelerating, leaders face a flood of information and uncertainty.

“When it comes to educating leaders about generative AI, it’s about helping them navigate the complexity of the space right now, because there’s so much hype and hundreds of papers published daily,” says Kraska. “The hard part is understanding which developments could actually have a chance of changing the field and which are just tiny improvements. There's a kind of FOMO [fear of missing out] for leaders that we can help reduce.”

Defining success: Shared goals for generative AI impact

Success within the initiative is defined by shared progress, open innovation, and mutual growth. “Consortium participants recognize, I think, that when I share my ideas with you, and you share your ideas with me, we’re both fundamentally better off,” explains Farias. “Progress on generative AI is not zero-sum, so it makes sense for this to be an open-source initiative.”

While participants may approach success from different angles, they share a common goal of advancing generative AI for broad societal benefit. “There will be many success metrics,” says Perakis. “We’ll educate students, who will be networking with companies. Companies will come together and learn from each other. Business leaders will come to MIT and have discussions that will help all of us, not just the leaders themselves.”

For Analog Devices’ Alan Lee, success is measured in tangible improvements that drive efficiency and product innovation: “For us at ADI, it’s a better, faster quality of experience for our customers, and that could mean better products. It could mean faster design cycles, faster verification cycles, and faster tuning of equipment that we already have or that we’re going to develop for the future. But beyond that, we want to help the world be a better, more efficient place.”

Ganesh highlights success through the lens of real-world application. “Success will also be defined by accelerating AI adoption within Tata companies, generating actionable knowledge that can be applied in real-world scenarios, and delivering significant advantages to our customers and stakeholders,” she says.

Generative AI is no longer confined to isolated research labs — it’s driving innovation across industries and disciplines. At MIT, the technology has become a campus-wide priority, connecting researchers, students, and industry leaders to solve complex challenges and uncover new opportunities. “It's truly an MIT initiative,” says Farias, “one that’s much larger than any individual or department on campus.”

The neural network artificial intelligence models used in applications like medical image processing and speech recognition perform operations on hugely complex data structures that require an enormous amount of computation to process. This is one reason deep-learning models consume so much energy.

To improve the efficiency of AI models, MIT researchers created an automated system that enables developers of deep learning algorithms to simultaneously take advantage of two types of data redundancy. This reduces the amount of computation, bandwidth, and memory storage needed for machine learning operations.

Existing techniques for optimizing algorithms can be cumbersome and typically only allow developers to capitalize on either sparsity or symmetry — two different types of redundancy that exist in deep learning data structures.

By enabling a developer to build an algorithm from scratch that takes advantage of both redundancies at once, the MIT researchers’ approach boosted the speed of computations by nearly 30 times in some experiments.

Because the system utilizes a user-friendly programming language, it could optimize machine-learning algorithms for a wide range of applications. The system could also help scientists who are not experts in deep learning but want to improve the efficiency of AI algorithms they use to process data. In addition, the system could have applications in scientific computing.

“For a long time, capturing these data redundancies has required a lot of implementation effort. Instead, a scientist can tell our system what they would like to compute in a more abstract way, without telling the system exactly how to compute it,” says Willow Ahrens, an MIT postdoc and co-author of a paper on the system, which will be presented at the International Symposium on Code Generation and Optimization.

She is joined on the paper by lead author Radha Patel ’23, SM ’24 and senior author Saman Amarasinghe, a professor in the Department of Electrical Engineering and Computer Science (EECS) and a principal researcher in the Computer Science and Artificial Intelligence Laboratory (CSAIL).

Cutting out computation

In machine learning, data are often represented and manipulated as multidimensional arrays known as tensors. A tensor is like a matrix, which is a rectangular array of values arranged on two axes, rows and columns. But unlike a two-dimensional matrix, a tensor can have many dimensions, or axes, making tensors more difficult to manipulate.

Deep-learning models perform operations on tensors using repeated matrix multiplication and addition — this process is how neural networks learn complex patterns in data. The sheer volume of calculations that must be performed on these multidimensional data structures requires an enormous amount of computation and energy.

But because of the way data in tensors are arranged, engineers can often boost the speed of a neural network by cutting out redundant computations.

For instance, if a tensor represents user review data from an e-commerce site, since not every user reviewed every product, most values in that tensor are likely zero. This type of data redundancy is called sparsity. A model can save time and computation by only storing and operating on non-zero values.

In addition, sometimes a tensor is symmetric, which means the top half and bottom half of the data structure are equal. In this case, the model only needs to operate on one half, reducing the amount of computation. This type of data redundancy is called symmetry.

“But when you try to capture both of these optimizations, the situation becomes quite complex,” Ahrens says.

To simplify the process, she and her collaborators built a new compiler, which is a computer program that translates complex code into a simpler language that can be processed by a machine. Their compiler, called SySTeC, can optimize computations by automatically taking advantage of both sparsity and symmetry in tensors.

They began the process of building SySTeC by identifying three key optimizations they can perform using symmetry.

First, if the algorithm’s output tensor is symmetric, then it only needs to compute one half of it. Second, if the input tensor is symmetric, then algorithm only needs to read one half of it. Finally, if intermediate results of tensor operations are symmetric, the algorithm can skip redundant computations.

Simultaneous optimizations

To use SySTeC, a developer inputs their program and the system automatically optimizes their code for all three types of symmetry. Then the second phase of SySTeC performs additional transformations to only store non-zero data values, optimizing the program for sparsity.

In the end, SySTeC generates ready-to-use code.

“In this way, we get the benefits of both optimizations. And the interesting thing about symmetry is, as your tensor has more dimensions, you can get even more savings on computation,” Ahrens says.

The researchers demonstrated speedups of nearly a factor of 30 with code generated automatically by SySTeC.

Because the system is automated, it could be especially useful in situations where a scientist wants to process data using an algorithm they are writing from scratch.

In the future, the researchers want to integrate SySTeC into existing sparse tensor compiler systems to create a seamless interface for users. In addition, they would like to use it to optimize code for more complicated programs.

This work is funded, in part, by Intel, the National Science Foundation, the Defense Advanced Research Projects Agency, and the Department of Energy.

Every cell in your body contains the same genetic sequence, yet each cell expresses only a subset of those genes. These cell-specific gene expression patterns, which ensure that a brain cell is different from a skin cell, are partly determined by the three-dimensional structure of the genetic material, which controls the accessibility of each gene.

MIT chemists have now come up with a new way to determine those 3D genome structures, using generative artificial intelligence. Their technique can predict thousands of structures in just minutes, making it much speedier than existing experimental methods for analyzing the structures.

Using this technique, researchers could more easily study how the 3D organization of the genome affects individual cells’ gene expression patterns and functions.

“Our goal was to try to predict the three-dimensional genome structure from the underlying DNA sequence,” says Bin Zhang, an associate professor of chemistry and the senior author of the study. “Now that we can do that, which puts this technique on par with the cutting-edge experimental techniques, it can really open up a lot of interesting opportunities.”

MIT graduate students Greg Schuette and Zhuohan Lao are the lead authors of the paper, which appears today in Science Advances.

From sequence to structure

Inside the cell nucleus, DNA and proteins form a complex called chromatin, which has several levels of organization, allowing cells to cram 2 meters of DNA into a nucleus that is only one-hundredth of a millimeter in diameter. Long strands of DNA wind around proteins called histones, giving rise to a structure somewhat like beads on a string.

Chemical tags known as epigenetic modifications can be attached to DNA at specific locations, and these tags, which vary by cell type, affect the folding of the chromatin and the accessibility of nearby genes. These differences in chromatin conformation help determine which genes are expressed in different cell types, or at different times within a given cell.

Over the past 20 years, scientists have developed experimental techniques for determining chromatin structures. One widely used technique, known as Hi-C, works by linking together neighboring DNA strands in the cell’s nucleus. Researchers can then determine which segments are located near each other by shredding the DNA into many tiny pieces and sequencing it.

This method can be used on large populations of cells to calculate an average structure for a section of chromatin, or on single cells to determine structures within that specific cell. However, Hi-C and similar techniques are labor-intensive, and it can take about a week to generate data from one cell.

To overcome those limitations, Zhang and his students developed a model that takes advantage of recent advances in generative AI to create a fast, accurate way to predict chromatin structures in single cells. The AI model that they designed can quickly analyze DNA sequences and predict the chromatin structures that those sequences might produce in a cell.

“Deep learning is really good at pattern recognition,” Zhang says. “It allows us to analyze very long DNA segments, thousands of base pairs, and figure out what is the important information encoded in those DNA base pairs.”

ChromoGen, the model that the researchers created, has two components. The first component, a deep learning model taught to “read” the genome, analyzes the information encoded in the underlying DNA sequence and chromatin accessibility data, the latter of which is widely available and cell type-specific.

The second component is a generative AI model that predicts physically accurate chromatin conformations, having been trained on more than 11 million chromatin conformations. These data were generated from experiments using Dip-C (a variant of Hi-C) on 16 cells from a line of human B lymphocytes.

When integrated, the first component informs the generative model how the cell type-specific environment influences the formation of different chromatin structures, and this scheme effectively captures sequence-structure relationships. For each sequence, the researchers use their model to generate many possible structures. That’s because DNA is a very disordered molecule, so a single DNA sequence can give rise to many different possible conformations.

“A major complicating factor of predicting the structure of the genome is that there isn’t a single solution that we’re aiming for. There’s a distribution of structures, no matter what portion of the genome you’re looking at. Predicting that very complicated, high-dimensional statistical distribution is something that is incredibly challenging to do,” Schuette says.

Rapid analysis

Once trained, the model can generate predictions on a much faster timescale than Hi-C or other experimental techniques.

“Whereas you might spend six months running experiments to get a few dozen structures in a given cell type, you can generate a thousand structures in a particular region with our model in 20 minutes on just one GPU,” Schuette says.

After training their model, the researchers used it to generate structure predictions for more than 2,000 DNA sequences, then compared them to the experimentally determined structures for those sequences. They found that the structures generated by the model were the same or very similar to those seen in the experimental data.

“We typically look at hundreds or thousands of conformations for each sequence, and that gives you a reasonable representation of the diversity of the structures that a particular region can have,” Zhang says. “If you repeat your experiment multiple times, in different cells, you will very likely end up with a very different conformation. That’s what our model is trying to predict.”

The researchers also found that the model could make accurate predictions for data from cell types other than the one it was trained on. This suggests that the model could be useful for analyzing how chromatin structures differ between cell types, and how those differences affect their function. The model could also be used to explore different chromatin states that can exist within a single cell, and how those changes affect gene expression.

“ChromoGen provides a new framework for AI-driven discovery of genome folding principles and demonstrates that generative AI can bridge genomic and epigenomic features with 3D genome structure, pointing to future work on studying the variation of genome structure and function across a broad range of biological contexts,” says Jian Ma, a professor of computational biology at Carnegie Mellon University, who was not involved in the research.

Another possible application would be to explore how mutations in a particular DNA sequence change the chromatin conformation, which could shed light on how such mutations may cause disease.

“There are a lot of interesting questions that I think we can address with this type of model,” Zhang says.

The researchers have made all of their data and the model available to others who wish to use it.

The research was funded by the National Institutes of Health.

Drone shows are an increasingly popular form of large-scale light display. These shows incorporate hundreds to thousands of airborne bots, each programmed to fly in paths that together form intricate shapes and patterns across the sky. When they go as planned, drone shows can be spectacular. But when one or more drones malfunction, as has happened recently in Florida, New York, and elsewhere, they can be a serious hazard to spectators on the ground.

Drone show accidents highlight the challenges of maintaining safety in what engineers call “multiagent systems” — systems of multiple coordinated, collaborative, and computer-programmed agents, such as robots, drones, and self-driving cars.

Now, a team of MIT engineers has developed a training method for multiagent systems that can guarantee their safe operation in crowded environments. The researchers found that once the method is used to train a small number of agents, the safety margins and controls learned by those agents can automatically scale to any larger number of agents, in a way that ensures the safety of the system as a whole.

In real-world demonstrations, the team trained a small number of palm-sized drones to safely carry out different objectives, from simultaneously switching positions midflight to landing on designated moving vehicles on the ground. In simulations, the researchers showed that the same programs, trained on a few drones, could be copied and scaled up to thousands of drones, enabling a large system of agents to safely accomplish the same tasks.

“This could be a standard for any application that requires a team of agents, such as warehouse robots, search-and-rescue drones, and self-driving cars,” says Chuchu Fan, associate professor of aeronautics and astronautics at MIT. “This provides a shield, or safety filter, saying each agent can continue with their mission, and we’ll tell you how to be safe.”

Fan and her colleagues report on their new method in a study appearing this month in the journal IEEE Transactions on Robotics. The study’s co-authors are MIT graduate students Songyuan Zhang and Oswin So as well as former MIT postdoc Kunal Garg, who is now an assistant professor at Arizona State University.

Mall margins

When engineers design for safety in any multiagent system, they typically have to consider the potential paths of every single agent with respect to every other agent in the system. This pair-wise path-planning is a time-consuming and computationally expensive process. And even then, safety is not guaranteed.

“In a drone show, each drone is given a specific trajectory — a set of waypoints and a set of times — and then they essentially close their eyes and follow the plan,” says Zhang, the study’s lead author. “Since they only know where they have to be and at what time, if there are unexpected things that happen, they don’t know how to adapt.”

The MIT team looked instead to develop a method to train a small number of agents to maneuver safely, in a way that could efficiently scale to any number of agents in the system. And, rather than plan specific paths for individual agents, the method would enable agents to continually map their safety margins, or boundaries beyond which they might be unsafe. An agent could then take any number of paths to accomplish its task, as long as it stays within its safety margins.

In some sense, the team says the method is similar to how humans intuitively navigate their surroundings.

“Say you’re in a really crowded shopping mall,” So explains. “You don’t care about anyone beyond the people who are in your immediate neighborhood, like the 5 meters surrounding you, in terms of getting around safely and not bumping into anyone. Our work takes a similar local approach.”

Safety barrier

In their new study, the team presents their method, GCBF+, which stands for “Graph Control Barrier Function.” A barrier function is a mathematical term used in robotics that calculates a sort of safety barrier, or a boundary beyond which an agent has a high probability of being unsafe. For any given agent, this safety zone can change moment to moment, as the agent moves among other agents that are themselves moving within the system.

When designers calculate barrier functions for any one agent in a multiagent system, they typically have to take into account the potential paths and interactions with every other agent in the system. Instead, the MIT team’s method calculates the safety zones of just a handful of agents, in a way that is accurate enough to represent the dynamics of many more agents in the system.

“Then we can sort of copy-paste this barrier function for every single agent, and then suddenly we have a graph of safety zones that works for any number of agents in the system,” So says.

To calculate an agent’s barrier function, the team’s method first takes into account an agent’s “sensing radius,” or how much of the surroundings an agent can observe, depending on its sensor capabilities. Just as in the shopping mall analogy, the researchers assume that the agent only cares about the agents that are within its sensing radius, in terms of keeping safe and avoiding collisions with those agents.

Then, using computer models that capture an agent’s particular mechanical capabilities and limits, the team simulates a “controller,” or a set of instructions for how the agent and a handful of similar agents should move around. They then run simulations of multiple agents moving along certain trajectories, and record whether and how they collide or otherwise interact.

“Once we have these trajectories, we can compute some laws that we want to minimize, like say, how many safety violations we have in the current controller,” Zhang says. “Then we update the controller to be safer.”

In this way, a controller can be programmed into actual agents, which would enable them to continually map their safety zone based on any other agents they can sense in their immediate surroundings, and then move within that safety zone to accomplish their task.

“Our controller is reactive,” Fan says. “We don’t preplan a path beforehand. Our controller is constantly taking in information about where an agent is going, what is its velocity, how fast other drones are going. It’s using all this information to come up with a plan on the fly and it’s replanning every time. So, if the situation changes, it’s always able to adapt to stay safe.”

The team demonstrated GCBF+ on a system of eight Crazyflies — lightweight, palm-sized quadrotor drones that they tasked with flying and switching positions in midair. If the drones were to do so by taking the straightest path, they would surely collide. But after training with the team’s method, the drones were able to make real-time adjustments to maneuver around each other, keeping within their respective safety zones, to successfully switch positions on the fly.

In similar fashion, the team tasked the drones with flying around, then landing on specific Turtlebots — wheeled robots with shell-like tops. The Turtlebots drove continuously around in a large circle, and the Crazyflies were able to avoid colliding with each other as they made their landings.

“Using our framework, we only need to give the drones their destinations instead of the whole collision-free trajectory, and the drones can figure out how to arrive at their destinations without collision themselves,” says Fan, who envisions the method could be applied to any multiagent system to guarantee its safety, including collision avoidance systems in drone shows, warehouse robots, autonomous driving vehicles, and drone delivery systems.

This work was partly supported by the U.S. National Science Foundation, MIT Lincoln Laboratory under the Safety in Aerobatic Flight Regimes (SAFR) program, and the Defence Science and Technology Agency of Singapore.

Highly energetic explosions in the sky are commonly attributed to gamma-ray bursts. We now understand that these bursts originate from either the merger of two neutron stars or the collapse of a massive star. In these scenarios, a newborn black hole is formed, emitting a jet that travels at nearly the speed of light. When these jets are directed toward Earth, we can observe them from vast distances — sometimes billions of light-years away — due to a relativistic effect known as Doppler boosting. Over the past decade, thousands of such gamma-ray bursts have been detected.

Since its launch in 2024, the Einstein Probe — an X-ray space telescope developed by the Chinese Academy of Sciences (CAS) in partnership with European Space Agency (ESA) and the Max Planck Institute for Extraterrestrial Physics — has been scanning the skies looking for energetic explosions, and in April the telescope observed an unusual event designated as EP240408A. Now an international team of astronomers, including Dheeraj Pasham from MIT, Igor Andreoni from University of North Carolina at Chapel Hill, and Brendan O’Connor from Carnegie Mellon University, and others have investigated this explosion using a slew of ground-based and space-based telescopes, including NuSTAR, Swift, Gemini, Keck, DECam, VLA, ATCA, and NICER, which was developed in collaboration with MIT. 

An open-access report of their findings, published Jan. 27 in The Astrophysical Journal Letters, indicates that the characteristics of this explosion do not match those of typical gamma-ray bursts. Instead, it may represent a rare new class of powerful cosmic explosion — a jetted tidal disruption event, which occurs when a supermassive black hole tears apart a star. 

“NICER’s ability to steer to pretty much any part of the sky and monitor for weeks has been instrumental in our understanding of these unusual cosmic explosions,” says Pasham, a research scientist at the MIT Kavli Institute for Astrophysics and Space Research.

While a jetted tidal disruption event is plausible, the researchers say the lack of radio emissions from this jet is puzzling. O’Connor surmises, “EP240408a ticks some of the boxes for several different kinds of phenomena, but it doesn’t tick all the boxes for anything. In particular, the short duration and high luminosity are hard to explain in other scenarios. The alternative is that we are seeing something entirely new!”

According to Pasham, the Einstein Probe is just beginning to scratch the surface of what seems possible. “I’m excited to chase the next weird explosion from the Einstein Probe”, he says, echoing astronomers worldwide who look forward to the prospect of discovering more unusual explosions from the farthest reaches of the cosmos.

Last year the Earth exceeded 1.5 degrees Celsius of warming above preindustrial times, a threshold beyond which wildfires, droughts, floods, and other climate impacts are expected to escalate in frequency, intensity, and lethality. To cap global warming at 1.5 C and avert that scenario, the nearly 200 signatory nations of the Paris Agreement on climate change will need to not only dramatically lower their greenhouse gas emissions, but also take measures to remove carbon dioxide (CO₂) from the atmosphere and durably store it at or below the Earth’s surface.

Past analyses of the climate mitigation potential, costs, benefits, and drawbacks of different carbon dioxide removal (CDR) options have focused primarily on three strategies: bioenergy with carbon capture and storage (BECCS), in which CO₂-absorbing plant matter is converted into fuels or directly burned to generate energy, with some of the plant’s carbon content captured and then stored safely and permanently; afforestation/reforestation, in which CO₂-absorbing trees are planted in large numbers; and direct air carbon capture and storage (DACCS), a technology that captures and separates CO₂ directly from ambient air, and injects it into geological reservoirs or incorporates it into durable products. 

To provide a more comprehensive and actionable analysis of CDR, a new study by researchers at the MIT Center for Sustainability Science and Strategy (CS3) first expands the option set to include biochar (charcoal produced from plant matter and stored in soil) and enhanced weathering (EW) (spreading finely ground rock particles on land to accelerate storage of CO₂ in soil and water). The study then evaluates portfolios of all five options — in isolation and in combination — to assess their capability to meet the 1.5 C goal, and their potential impacts on land, energy, and policy costs.

The study appears in the journal Environmental Research Letters. Aided by their global multi-region, multi-sector Economic Projection and Policy Analysis (EPPA) model, the MIT CS3 researchers produce three key findings.

First, the most cost-effective, low-impact strategy that policymakers can take to achieve global net-zero emissions — an essential step in meeting the 1.5 C goal — is to diversify their CDR portfolio, rather than rely on any single option. This approach minimizes overall cropland and energy consumption, and negative impacts such as increased food insecurity and decreased energy supplies.

By diversifying across multiple CDR options, the highest CDR deployment of around 31.5 gigatons of CO₂ per year is achieved in 2100, while also proving the most cost-effective net-zero strategy. The study identifies BECCS and biochar as most cost-competitive in removing CO₂ from the atmosphere, followed by EW, with DACCS as uncompetitive due to high capital and energy requirements. While posing logistical and other challenges, biochar and EW have the potential to improve soil quality and productivity across 45 percent of all croplands by 2100.

“Diversifying CDR portfolios is the most cost-effective net-zero strategy because it avoids relying on a single CDR option, thereby reducing and redistributing negative impacts on agriculture, forestry, and other land uses, as well as on the energy sector,” says Solene Chiquier, lead author of the study who was a CS3 postdoc during its preparation.

The second finding: There is no optimal CDR portfolio that will work well at global and national levels. The ideal CDR portfolio for a particular region will depend on local technological, economic, and geophysical conditions. For example, afforestation and reforestation would be of great benefit in places like Brazil, Latin America, and Africa, by not only sequestering carbon in more acreage of protected forest but also helping to preserve planetary well-being and human health.

“In designing a sustainable, cost-effective CDR portfolio, it is important to account for regional availability of agricultural, energy, and carbon-storage resources,” says Sergey Paltsev, CS3 deputy director, MIT Energy Initiative senior research scientist, and supervising co-author of the study. “Our study highlights the need for enhancing knowledge about local conditions that favor some CDR options over others.”

Finally, the MIT CS3 researchers show that delaying large-scale deployment of CDR portfolios could be very costly, leading to considerably higher carbon prices across the globe — a development sure to deter the climate mitigation efforts needed to achieve the 1.5 C goal. They recommend near-term implementation of policy and financial incentives to help fast-track those efforts.

A home robot trained to perform household tasks in a factory may fail to effectively scrub the sink or take out the trash when deployed in a user’s kitchen, since this new environment differs from its training space.

To avoid this, engineers often try to match the simulated training environment as closely as possible with the real world where the agent will be deployed.

However, researchers from MIT and elsewhere have now found that, despite this conventional wisdom, sometimes training in a completely different environment yields a better-performing artificial intelligence agent.

Their results indicate that, in some situations, training a simulated AI agent in a world with less uncertainty, or “noise,” enabled it to perform better than a competing AI agent trained in the same, noisy world they used to test both agents.

The researchers call this unexpected phenomenon the indoor training effect.

“If we learn to play tennis in an indoor environment where there is no noise, we might be able to more easily master different shots. Then, if we move to a noisier environment, like a windy tennis court, we could have a higher probability of playing tennis well than if we started learning in the windy environment,” explains Serena Bono, a research assistant in the MIT Media Lab and lead author of a paper on the indoor training effect.

The researchers studied this phenomenon by training AI agents to play Atari games, which they modified by adding some unpredictability. They were surprised to find that the indoor training effect consistently occurred across Atari games and game variations.

They hope these results fuel additional research toward developing better training methods for AI agents.

“This is an entirely new axis to think about. Rather than trying to match the training and testing environments, we may be able to construct simulated environments where an AI agent learns even better,” adds co-author Spandan Madan, a graduate student at Harvard University.

Bono and Madan are joined on the paper by Ishaan Grover, an MIT graduate student; Mao Yasueda, a graduate student at Yale University; Cynthia Breazeal, professor of media arts and sciences and leader of the Personal Robotics Group in the MIT Media Lab; Hanspeter Pfister, the An Wang Professor of Computer Science at Harvard; and Gabriel Kreiman, a professor at Harvard Medical School. The research will be presented at the Association for the Advancement of Artificial Intelligence Conference.

Training troubles

The researchers set out to explore why reinforcement learning agents tend to have such dismal performance when tested on environments that differ from their training space.

Reinforcement learning is a trial-and-error method in which the agent explores a training space and learns to take actions that maximize its reward.

The team developed a technique to explicitly add a certain amount of noise to one element of the reinforcement learning problem called the transition function. The transition function defines the probability an agent will move from one state to another, based on the action it chooses.

If the agent is playing Pac-Man, a transition function might define the probability that ghosts on the game board will move up, down, left, or right. In standard reinforcement learning, the AI would be trained and tested using the same transition function.

The researchers added noise to the transition function with this conventional approach and, as expected, it hurt the agent’s Pac-Man performance.

But when the researchers trained the agent with a noise-free Pac-Man game, then tested it in an environment where they injected noise into the transition function, it performed better than an agent trained on the noisy game.

“The rule of thumb is that you should try to capture the deployment condition’s transition function as well as you can during training to get the most bang for your buck. We really tested this insight to death because we couldn’t believe it ourselves,” Madan says.

Injecting varying amounts of noise into the transition function let the researchers test many environments, but it didn’t create realistic games. The more noise they injected into Pac-Man, the more likely ghosts would randomly teleport to different squares.

To see if the indoor training effect occurred in normal Pac-Man games, they adjusted underlying probabilities so ghosts moved normally but were more likely to move up and down, rather than left and right. AI agents trained in noise-free environments still performed better in these realistic games.

“It was not only due to the way we added noise to create ad hoc environments. This seems to be a property of the reinforcement learning problem. And that was even more surprising to see,” Bono says.

Exploration explanations

When the researchers dug deeper in search of an explanation, they saw some correlations in how the AI agents explore the training space.

When both AI agents explore mostly the same areas, the agent trained in the non-noisy environment performs better, perhaps because it is easier for the agent to learn the rules of the game without the interference of noise.

If their exploration patterns are different, then the agent trained in the noisy environment tends to perform better. This might occur because the agent needs to understand patterns it can’t learn in the noise-free environment.

“If I only learn to play tennis with my forehand in the non-noisy environment, but then in the noisy one I have to also play with my backhand, I won’t play as well in the non-noisy environment,” Bono explains.

In the future, the researchers hope to explore how the indoor training effect might occur in more complex reinforcement learning environments, or with other techniques like computer vision and natural language processing. They also want to build training environments designed to leverage the indoor training effect, which could help AI agents perform better in uncertain environments.

In biology textbooks, the endoplasmic reticulum is often portrayed as a distinct, compact organelle near the nucleus, and is commonly known to be responsible for protein trafficking and secretion. In reality, the ER is vast and dynamic, spread throughout the cell and able to establish contact and communication with and between other organelles. These membrane contacts regulate processes as diverse as fat metabolism, sugar metabolism, and immune responses.

Exploring how pathogens manipulate and hijack essential processes to promote their own life cycles can reveal much about fundamental cellular functions and provide insight into viable treatment options for understudied pathogens.

New research from the Lamason Lab in the Department of Biology at MIT recently published in the Journal of Cell Biology has shown that Rickettsia parkeri, a bacterial pathogen that lives freely in the cytosol, can interact in an extensive and stable way with the rough endoplasmic reticulum, forming previously unseen contacts with the organelle.

It’s the first known example of a direct interkingdom contact site between an intracellular bacterial pathogen and a eukaryotic membrane.

The Lamason Lab studies R. parkeri as a model for infection of the more virulent Rickettsia rickettsii. R. rickettsii, carried and transmitted by ticks, causes Rocky Mountain Spotted Fever. Left untreated, the infection can cause symptoms as severe as organ failure and death.

Rickettsia is difficult to study because it is an obligate pathogen, meaning it can only live and reproduce inside living cells, much like a virus. Researchers must get creative to parse out fundamental questions and molecular players in the R. parkeri life cycle, and much remains unclear about how R. parkeri spreads.

Detour to the junction

First author Yamilex Acevedo-Sánchez, a BSG-MSRP-Bio program alum and a graduate student at the time, stumbled across the ER and R. parkeri interactions while trying to observe Rickettsia reaching a cell junction.

The current model for Rickettsia infection involves R. parkeri spreading cell to cell by traveling to the specialized contact sites between cells and being engulfed by the neighboring cell in order to spread. Listeria monocytogenes, which the Lamason Lab also studies, uses actin tails to forcefully propel itself into a neighboring cell. By contrast, R. parkeri can form an actin tail, but loses it before reaching the cell junction. Somehow, R. parkeri is still able to spread to neighboring cells.

After an MIT seminar about the ER’s lesser-known functions, Acevedo-Sánchez developed a cell line to observe whether Rickettsia might be spreading to neighboring cells by hitching a ride on the ER to reach the cell junction.

Instead, she saw an unexpectedly high percentage of R. parkeri surrounded and enveloped by the ER, at a distance of about 55 nanometers. This distance is significant because membrane contacts for interorganelle communication in eukaryotic cells form connections from 10-80 nanometers wide. The researchers ruled out that what they saw was not an immune response, and the sections of the ER interacting with the R. parkeri were still connected to the wider network of the ER.

“I’m of the mind that if you want to learn new biology, just look at cells,” Acevedo-Sánchez says. “Manipulating the organelle that establishes contact with other organelles could be a great way for a pathogen to gain control during infection.” 

The stable connections were unexpected because the ER is constantly breaking and reforming connections, lasting seconds or minutes. It was surprising to see the ER stably associating around the bacteria. As a cytosolic pathogen that exists freely in the cytosol of the cells it infects, it was also unexpected to see R. parkeri surrounded by a membrane at all.

Small margins

Acevedo-Sánchez collaborated with the Center for Nanoscale Systems at Harvard University to view her initial observations at higher resolution using focused ion beam scanning electron microscopy. FIB-SEM involves taking a sample of cells and blasting them with a focused ion beam in order to shave off a section of the block of cells. With each layer, a high-resolution image is taken. The result of this process is a stack of images.

From there, Acevedo-Sánchez marked what different areas of the images were — such as the mitochondria, Rickettsia, or the ER — and a program called ORS Dragonfly, a machine learning program, sorted through the thousand or so images to identify those categories. That information was then used to create 3D models of the samples. 

Acevedo-Sánchez noted that less than 5 percent of R. parkeri formed connections with the ER — but small quantities of certain characteristics are known to be critical for R. parkeri infection. R. parkeri can exist in two states: motile, with an actin tail, and nonmotile, without it. In mutants unable to form actin tails, R. parkeri are unable to progress to adjacent cells — but in nonmutants, the percentage of R. parkeri that have tails starts at about 2 percent in early infection and never exceeds 15 percent at the height of it.

The ER only interacts with nonmotile R. parkeri, and those interactions increased 25-fold in mutants that couldn’t form tails.

Creating connections

Co-authors Acevedo-Sánchez, Patrick Woida, and Caroline Anderson also investigated possible ways the connections with the ER are mediated. VAP proteins, which mediate ER interactions with other organelles, are known to be co-opted by other pathogens during infection.

During infection by R. parkeri, VAP proteins were recruited to the bacteria; when VAP proteins were knocked out, the frequency of interactions between R. parkeri and the ER decreased, indicating R. parkeri may be taking advantage of these cellular mechanisms for its own purposes during infection.

Although Acevedo-Sánchez now works as a senior scientist at AbbVie, the Lamason Lab is continuing the work of exploring the molecular players that may be involved, how these interactions are mediated, and whether the contacts affect the host or bacteria’s life cycle.

Senior author and associate professor of biology Rebecca Lamason noted that these potential interactions are particularly interesting because bacteria and mitochondria are thought to have evolved from a common ancestor. The Lamason Lab has been exploring whether R. parkeri could form the same membrane contacts that mitochondria do, although they haven’t proven that yet. So far, R. parkeri is the only cytosolic pathogen that has been observed behaving this way.

“It’s not just bacteria accidentally bumping into the ER. These interactions are extremely stable. The ER is clearly extensively wrapping around the bacterium, and is still connected to the ER network,” Lamason says. “It seems like it has a purpose — what that purpose is remains a mystery.” 

“There is no treatment available for your son. We can’t do anything to help him.”

When Fernando Goldsztein MBA ’03 heard those words, something inside him snapped.

“I refused to accept what the doctors were saying. I transformed my fear into my greatest strength and started fighting.”

Goldsztein’s 12-year-old son Frederico was diagnosed with relapsing medulloblastoma, a life-threatening pediatric brain tumor. Goldsztein's life — and career plan — changed in an instant. He had to learn to become a different kind of leader altogether.

While Goldsztein never set out to become a founder, the MIT Sloan School of Management taught him the importance of networking, building friendships, and making career connections with peers and faculty from all walks of life. He began using those skills in a new way — boldly reaching out to the top medulloblastoma doctors and scientists at hospitals around the world to ask for help.

“I knew that I had to do something to save Frederico, but also the other estimated 15,000 children diagnosed with the disease around the world each year,” he says.

In 2021, Goldsztein launched The Medulloblastoma Initiative (MBI), a nonprofit organization dedicated to finding a cure using a remarkable new model for funding rare disease research.

In just 18 months, the organization — which is still in startup mode — has raised $11 million in private funding and brought together 14 of the world’s most prestigious labs and hospitals from across North America, Europe, and Brazil.

Two promising trials will launch in the coming months, and three additional trials are in the pipeline and currently awaiting U.S. Food and Drug Administration approval.

All of this in an industry that is notorious for bureaucratic red tape, and where the timeline from an initial lab discovery to a patient receiving a first treatment averages seven to 15 years.

While government research grants typically allocate just 4 cents on the dollar toward pediatric cancer research — pennies doled out across multiple labs pursuing uncoordinated efforts — MBI is laser-focused on pushing 100 percent of their funding toward a singular goal, without any overhead or administrative costs.

“There is no time to lose,” Goldsztein says. “We are making science move faster than it ever has before.”

The MBI blueprint for funding cures for rare diseases is replicable, and likely to disrupt the standard way health care research is funded and carried out by radically shortening the timeline.

From despair to strength

After his initial diagnosis at age 9, Frederico went through a nine-hour brain surgery and came to the United States to receive standard treatment. Goldsztein looked on helplessly as his son received radiation and then nine grueling rounds of chemotherapy.

First pioneered in the 1980s, this standard treatment protocol cures 70 percent of children. Still, it leaves most of them with lifelong side effects like cognitive problems, endocrine issues that stunt growth, and secondary tumors. Frederico was on the wrong side of that statistic. Just three years later, his tumor relapsed.

Goldsztein grimaces as he recalls the prognosis he and his wife heard from the doctors.

“It was unbelievable to me that there had been almost no discoveries in 40 years,” he says.

Ultimately, he found hope and partnership in Roger Packer, the director of the Brain Tumor Institute and the Gilbert Family Neurofibromatosis Institute of Children’s National Hospital. He is also the very doctor who created the standard treatment years before.

Packer explains that finding effective therapies for medulloblastoma was complex for 30 years because it is an umbrella term for 13 types of tumors. Frederico suffers from the most common one, Group 4.

Part of the reason the treatment has not changed is that, until recently, medicine has not advanced enough to detect differences between the different tumor types. Packer explains, “Now with molecular genetic testing and methylation, which is a way to essentially sort tumors, that has changed.”

The problem for Frederico was that very few researchers were working on Group 4, the sub-type of medulloblastoma that is the most common tumor, yet also the one that scientists know the least about.

Goldsztein challenged Packer: “If I can get you the funding, what can your lab do to advance medulloblastoma research quickly?”

An open-source consortium model

Packer advised that they work together to “try something different,” instead of just throwing money at research without any guideposts.

“We set up a consortium of leading institutions around the world doing medulloblastoma research, asked them to change their lab approach to focus on the Group 4 tumor, and assigned each lab a question to answer. We charged them with coming up with therapy — not in seven to 10 years, which is the normal transition from discovery to developing a drug and getting it to a patient, but within a two-year timeline,” he says.

Initially, seven labs signed on. Today, the Cure Group 4 Consortium is made up of 14 partners and reads like a who’s who of medulloblastoma heavy hitters: Children’s National Hospital, SickKids, Hopp Children’s Cancer Center, and Texas Children’s Hospital.

Labs can only join the consortium if they agree to follow some unusual rules. As Goldsztein explains, “To be accepted into this group and receive funding, there are no silos, and there is no duplicated work. Everyone has a piece of the puzzle, and we work together to move fast. That is the magic of our model.”

Inspired by MIT’s open-source methods, researchers must share data freely with one another to accelerate the group’s overall progress. This kind of partnership across labs and borders is unprecedented in a highly competitive sector.

Mariano Gargiulo MBA ’03 met Goldsztein on the first day of their MIT Sloan Fellows MBA program orientation and has been his dear friend ever since. An early-stage donor to MBI and a Houston-based executive in the energy sector, Gargiulo sat down with Goldsztein as he first conceptualized MBI’s operating model.

“Usually, startup business models plot out the next 10-15 years; Fernando’s timeline was only two years, and his benchmarks were in three-month increments.” It was audaciously optimistic, says Gargiulo, but so was the founder.

“When I saw it, I did not doubt that he would achieve his goals. I’m seeing Fernando hit those first targets now and it’s amazing to watch,” Gargiulo says.

Children’s National Hospital endorsed MBI in 2023 and invited Goldsztein to sit on its foundation’s board, adding credibility to the initiative and his ability to fundraise more ambitiously.

According to Packer, in the next few months, the first two MBI protocols will reach patients for the first time: an immunotherapy protocol, which “leverages the body’s immune response to target cancer cells more effectively and safely than traditional therapies,” and a medulloblastoma vaccine, which “adapts similar methodologies used in Covid-19 vaccine development. This approach aims to provide a versatile and mobile treatment that could be distributed globally.”

A matter of when

When Goldsztein is not with his own family in Brazil, fundraising, or managing MBI, he is on Zoom with a network of more than 70 other families with children with relapsed medulloblastoma. “I’m not a doctor and I don’t give out medical advice, but with these trials, we are giving each other hope,” he explains.

Hope and purpose are commodities that Goldsztein has in spades. “I don’t understand the idea of doing business and accumulating assets, but not helping others,” he says. He shared that message with an auditorium of his fellow alumni at his 2023 MIT Sloan Reunion.

Frederico, who defied all odds and lived with the threat of recurrence, recently graduated high school. He is interested in international relations and passionate about photography. “This is about finding a cure for Frederico and for all kids,” Goldsztein says.

When asked how the world would be impacted if MBI found a cure for medulloblastoma, Goldsztein shakes his head.

“We are going to find the cure. It’s not if, it’s a matter of when.”

His next goal is to scale MBI and have it serve as a resource for groups that want to replicate its playbook to solve other childhood diseases.

“I’m never going to stop,” he says.

Buildings cost a lot these days. But when concrete buildings are being constructed, there’s another material that can make them less expensive: mud.

MIT researchers have developed a method to use lightly treated mud, including soil from a building site, as the “formwork” molds into which concrete is poured. The technique deploys 3D printing and can replace the more costly method of building elaborate wood formworks for concrete construction.

“What we’ve demonstrated is that we can essentially take the ground we’re standing on, or waste soil from a construction site, and transform it into accurate, highly complex, and flexible formwork for customized concrete structures,” says Sandy Curth, a PhD student in MIT’s Department of Architecture who has helped spearhead the project.

The approach could help concrete-based construction take place more quickly and efficiently. It could also reduce costs and carbon emissions.

“It has the potential for immediate impact and doesn’t require changing the nature of the construction industry,” says Curth, who doubles as director of the Programmable Mud Initiative.

Curth has co-authored multiple papers about the method, most recently, “EarthWorks: Zero waste 3D printed earthen formwork for shape-optimized, reinforced concrete construction,” published in the journal Construction and Building Materials. Curth wrote that paper with nine co-authors, including Natalie Pearl, Emily Wissemann, Tim Cousin, Latifa Alkhayat, Vincent Jackow, Keith Lee, and Oliver Moldow, all MIT students; and Mohamed Ismail of the University of Virginia.

The paper’s final two co-authors are Lawrence Sass, professor and chair of the Computation Group in MIT’s Department of Architecture, and Caitlin Mueller, an associate professor at MIT in the Department of Architecture and the Department of Civil and Environmental Engineering. Sass is Curth’s graduate advisor.

Building a structure once, not twice

Constructing wooden formwork for a building is costly and time-consuming. There is saying in the industry that concrete structures have to be built twice — once through the wooden formwork, then again in the concrete poured into the forms.

Using soil for the formwork could change that process. While it might seem like an unusual material compared to the solidity of wooden formwork, soil is firm enough to handle poured concrete. The EarthWorks method, as its known, introduces some additive materials, such as straw, and a wax-like coating for the soil material to prevent any water from draining out of the concrete. Using large-scale 3D printing, the researchers can take soil from a construction site and print it into a custom-designed formwork shape.

“What we’ve done is make a system where we are using what is largely straightforward, large-scale 3D printing technology, and making it highly functional for the material,” Curth says. “We found a way to make formwork that is infinitely recyclable. It’s just dirt.”

Beyond cost and ease of acquiring the materials, the method offers at least two other interrelated advantages. One is environmental: Concrete construction accounts for as much as 8 percent of global carbon emissions, and this approach supports substantial emissions reductions, both through the formwork material itself and the ease of shaping the resulting concrete to only use what is structurally required. Using a method called shape optimization, developed for reinforced concrete in previous research by Ismail and Mueller, it is possible to reduce the carbon emissions of concrete structural frames by more than 50 percent.  

“The EarthWorks technique brings these complex, optimized structures much closer to built reality by offering a low-cost, low-carbon fabrication technique for formwork that can be deployed anywhere in the world,” Mueller says.

“It’s an enabling technology to make reinforced concrete buildings much, much more materially efficient, which has a direct impact on global carbon emissions,” Curth adds.

More generally, the EarthWorks method allows architects and engineers to create customized concrete shapes more easily, due to the flexibility of the formwork material. It is easier to cast concrete in an unusual shape when molding it with soil, not wood.

“What’s cool here is we’re able to make shape-optimized building elements for the same amount of time and energy it would take to make rectilinear building elements,” Curth says.

Group project

As Curth notes, the projects developed by the Programmable Mud group are highly collaborative. He emphasizes the roles played by both Sass, a leader in using computation to help develop low-cost housing, and Mueller, whose work also deploys new computational methods to assess innovative structural ideas in architecture.

“Concrete is a wonderful material when it is used thoughtfully and efficiently, which is inherently connected to how it is shaped,” Mueller says. “However, the minimal forms that emerge from optimization are at odds with conventional construction logics. It is very exciting to advance a technique that subverts this supposed tradeoff, showing that performance-driven complexity can be achieved with low carbon emissions and low cost.”

While finishing his doctorate at MIT, Curth has also founded a firm, FORMA Systems, through which he hopes to take the EarthWorks method into the construction industry. Using this approach does mean builders would need to have a large 3D printer on-site. However, they would also save significantly on materials costs, he says.

Further in the future, Curth envisions a time when the method could be used not just for formworks, but to construct templates for, say, two-story residential building made entirely out of earth. Of course, some parts of the world, including the U.S., extensively use adobe architecture already, but the idea here would be to systematize the production of such homes and make them inexpensive in the process.

In either case, Curth says, as formwork for concrete or by itself, we now have new ways to apply soil to construction.

“People have built with earth for as long as we’ve had buildings, but given contemporary demands for urban concrete buildings, this approach basically decouples cost from complexity,” Curth says. “I guarantee you we can start to make higher-performance buildings for less money.”

The project  was supported by the Sidara Urban Research Seed Fund administered by MIT’s Leventhal Center for Advanced Urbanism, and by lyndaLABS.

Several years ago, the residents of a manufactured-home neighborhood in southeast suburban Houston, not far from the Buffalo Bayou, took a major step in dealing with climate problems: They bought the land under their homes. Then they installed better drainage and developed strategies to share expertise and tools for home repairs. The result? The neighborhood made it through Hurricane Harvey in 2017 and a winter freeze in 2021 without major damage.

The neighborhood is part of a U.S. movement toward the Resident Owned Community (ROC) model for manufactured home parks. Many people in manufactured homes — mobile homes — do not own the land under them. But if the residents of a manufactured-home park can form an ROC, they can take action to adapt to climate risks — and ease the threat of eviction. With an ROC, manufactured-home residents can be there to stay.

That speaks to a larger issue: In cities, lower-income residents are often especially vulnerable to natural hazards, such as flooding, extreme heat, and wildfire. But efforts aimed at helping cities as a whole withstand these disasters can lead to interventions that displace already-disadvantaged residents — by turning a low-lying neighborhood into a storm buffer, for instance.

“The global climate crisis has very differential effects on cities, and neighborhoods within cities,” says Lawrence Vale, a professor of urban studies at MIT and co-author of a new book on the subject, “The Equitably Resilient City,” published by the MIT Press and co-authored with Zachary B. Lamb PhD ’18, an assistant professor at the University of California at Berkeley.

In the book, the scholars delve into 12 case studies from around the globe which, they believe, have it both ways: Low- and middle-income communities have driven climate progress through tangible built projects, while also keeping people from being displaced, and indeed helping them participate in local governance and neighborhood decision-making.

“We can either dive into despair about climate issues, or think they’re solvable and ask what it takes to succeed in a more equitable way,” says Vale, who is the Ford Professor of Urban Design and Planning at MIT. “This book is asking how people look at problems more holistically — to show how environmental impacts are integrated with their livelihoods, with feeling they can have security from displacement, and feeling they’re not going to be displaced, with being empowered to share in the governance where they live.”

As Lamb notes, “Pursuing equitable urban climate adaptation requires both changes in the physical built environment of cities and innovations in institutions and governance practices to address deep-seated causes of inequality.”

Twelve projects, four elements

Research for “The Equitably Resilient City” began with exploration of about 200 potential cases, and ultimately focused on 12 projects from around the globe, including the U.S., Brazil, Thailand, and France. Vale and Lamb, coordinating with locally-based research teams, visited these diverse sites and conducted interviews in nine languages.

All 12 projects work on multiple levels at once: They are steps toward environmental progress that also help local communities in civic and economic terms. The book uses the acronym LEGS (“livelihood, environment, governance, and security”) to encapsulate this need to make equitable progress on four different fronts.

“Doing one of those things well is worth recognition, and doing all of them well is exciting,” Vale says. “It’s important to understand not just what these communities did, but how they did it and whose views were involved. These 12 cases are not a random sample. The book looks for people who are partially succeeding at difficult things in difficult circumstances.”

One case study is set in São Paolo, Brazil, where low-income residents of a hilly favela benefitted from new housing in the area on undeveloped land that is less prone to slides. In San Juan, Puerto Rico, residents of low-lying neighborhoods abutting a water channel formed a durable set of community groups to create a fairer solution to flooding: Although the channel needed to be re-widened, the local coalition insisted on limiting displacement, supporting local livelihoods and improving environmental conditions and public space.

“There is a backlash to older practices,” Vale says, referring to the large-scale urban planning and infrastructure projects of the mid-20^th century, which often ignored community input. “People saw what happened during the urban renewal era and said, ‘You’re not going to do that to us again.’”

Indeed, one through-line in “The Equitably Resilient City” is that cities, like all places, can be contested political terrain. Often, solid solutions emerge when local groups organize, advocate for new solutions, and eventually gain enough traction to enact them.

“Every one of our examples and cases has probably 15 or 20 years of activity behind it, as well as engagements with a much deeper history,” Vale says. “They’re all rooted in a very often troubled [political] context. And yet these are places that have made progress possible.”

Think locally, adapt anywhere

Another motif of “The Equitably Resilient City” is that local progress matters greatly, for a few reasons — including the value of having communities develop projects that meet their own needs, based on their input. Vale and Lamb are interested in projects even if they are very small-scale, and devote one chapter of the book to the Paris OASIS program, which has developed a series of cleverly designed, heavily tree-dotted school playgrounds across Paris. These projects provide environmental education opportunities and help mitigate flooding and urban heat while adding CO2-harnessing greenery to the cityscape.

An individual park, by itself, can only do so much, but the concept behind it can be adopted by anyone.

“This book is mostly centered on local projects rather than national schemes,” Vale says. “The hope is they serve as an inspiration for people to adapt to their own situations.”

After all, the urban geography and governance of places such as Paris or São Paulo will differ widely. But efforts to make improvements to public open space or to well-located inexpensive housing stock applies in cities across the world.

Similarly, the authors devote a chapter to work in the Cully neighborhood in Portland, Oregon, where community leaders have instituted a raft of urban environmental improvements while creating and preserving more affordable housing. The idea in the Cully area, as in all these cases, is to make places more resistant to climate change while enhancing them as good places to live for those already there.

“Climate adaptation is going to mobilize enormous public and private resources to reshape cities across the globe,” Lamb notes. “These cases suggest pathways where those resources can make cities both more resilient in the face of climate change and more equitable. In fact, these projects show how making cities more equitable can be part of making them more resilient.”

Other scholars have praised the book. Eric Klinenberg, director of New York University’s Institute for Public Knowledge has called it “at once scholarly, constructive, and uplifting, a reminder that better, more just cities remain within our reach.”

Vale also teaches some of the book’s concepts in his classes, finding that MIT students, wherever they are from, enjoy the idea of thinking creatively about climate resilience.

“At MIT, students want to find ways of applying technical skills to urgent global challenges,” Vale says. “I do think there are many opportunities, especially at a time of climate crisis. We try to highlight some of the solutions that are out there. Give us an opportunity, and we’ll show you what a place can be.”

As the capabilities of generative AI models have grown, you've probably seen how they can transform simple text prompts into hyperrealistic images and even extended video clips.

More recently, generative AI has shown potential in helping chemists and biologists explore static molecules, like proteins and DNA. Models like AlphaFold can predict molecular structures to accelerate drug discovery, and the MIT-assisted “RFdiffusion,” for example, can help design new proteins. One challenge, though, is that molecules are constantly moving and jiggling, which is important to model when constructing new proteins and drugs. Simulating these motions on a computer using physics — a technique known as molecular dynamics — can be very expensive, requiring billions of time steps on supercomputers.

As a step toward simulating these behaviors more efficiently, MIT Computer Science and Artificial Intelligence Laboratory (CSAIL) and Department of Mathematics researchers have developed a generative model that learns from prior data. The team’s system, called MDGen, can take a frame of a 3D molecule and simulate what will happen next like a video, connect separate stills, and even fill in missing frames. By hitting the “play button” on molecules, the tool could potentially help chemists design new molecules and closely study how well their drug prototypes for cancer and other diseases would interact with the molecular structure it intends to impact.

Co-lead author Bowen Jing SM ’22 says that MDGen is an early proof of concept, but it suggests the beginning of an exciting new research direction. “Early on, generative AI models produced somewhat simple videos, like a person blinking or a dog wagging its tail,” says Jing, a PhD student at CSAIL. “Fast forward a few years, and now we have amazing models like Sora or Veo that can be useful in all sorts of interesting ways. We hope to instill a similar vision for the molecular world, where dynamics trajectories are the videos. For example, you can give the model the first and 10th frame, and it’ll animate what’s in between, or it can remove noise from a molecular video and guess what was hidden.”

The researchers say that MDGen represents a paradigm shift from previous comparable works with generative AI in a way that enables much broader use cases. Previous approaches were “autoregressive,” meaning they relied on the previous still frame to build the next, starting from the very first frame to create a video sequence. In contrast, MDGen generates the frames in parallel with diffusion. This means MDGen can be used to, for example, connect frames at the endpoints, or “upsample” a low frame-rate trajectory in addition to pressing play on the initial frame.

This work was presented in a paper shown at the Conference on Neural Information Processing Systems (NeurIPS) this past December. Last summer, it was awarded for its potential commercial impact at the International Conference on Machine Learning’s ML4LMS Workshop.

Some small steps forward for molecular dynamics

In experiments, Jing and his colleagues found that MDGen’s simulations were similar to running the physical simulations directly, while producing trajectories 10 to 100 times faster.

The team first tested their model’s ability to take in a 3D frame of a molecule and generate the next 100 nanoseconds. Their system pieced together successive 10-nanosecond blocks for these generations to reach that duration. The team found that MDGen was able to compete with the accuracy of a baseline model, while completing the video generation process in roughly a minute — a mere fraction of the three hours that it took the baseline model to simulate the same dynamic.

When given the first and last frame of a one-nanosecond sequence, MDGen also modeled the steps in between. The researchers’ system demonstrated a degree of realism in over 100,000 different predictions: It simulated more likely molecular trajectories than its baselines on clips shorter than 100 nanoseconds. In these tests, MDGen also indicated an ability to generalize on peptides it hadn’t seen before.

MDGen’s capabilities also include simulating frames within frames, “upsampling” the steps between each nanosecond to capture faster molecular phenomena more adequately. It can even ​​“inpaint” structures of molecules, restoring information about them that was removed. These features could eventually be used by researchers to design proteins based on a specification of how different parts of the molecule should move.

Toying around with protein dynamics

Jing and co-lead author Hannes Stärk say that MDGen is an early sign of progress toward generating molecular dynamics more efficiently. Still, they lack the data to make these models immediately impactful in designing drugs or molecules that induce the movements chemists will want to see in a target structure.

The researchers aim to scale MDGen from modeling molecules to predicting how proteins will change over time. “Currently, we’re using toy systems,” says Stärk, also a PhD student at CSAIL. “To enhance MDGen’s predictive capabilities to model proteins, we’ll need to build on the current architecture and data available. We don’t have a YouTube-scale repository for those types of simulations yet, so we’re hoping to develop a separate machine-learning method that can speed up the data collection process for our model.”

For now, MDGen presents an encouraging path forward in modeling molecular changes invisible to the naked eye. Chemists could also use these simulations to delve deeper into the behavior of medicine prototypes for diseases like cancer or tuberculosis.

“Machine learning methods that learn from physical simulation represent a burgeoning new frontier in AI for science,” says Bonnie Berger, MIT Simons Professor of Mathematics, CSAIL principal investigator, and senior author on the paper. “MDGen is a versatile, multipurpose modeling framework that connects these two domains, and we’re very excited to share our early models in this direction.”

“Sampling realistic transition paths between molecular states is a major challenge,” says fellow senior author Tommi Jaakkola, who is the MIT Thomas Siebel Professor of electrical engineering and computer science and the Institute for Data, Systems, and Society, and a CSAIL principal investigator. “This early work shows how we might begin to address such challenges by shifting generative modeling to full simulation runs.”

Researchers across the field of bioinformatics have heralded this system for its ability to simulate molecular transformations. “MDGen models molecular dynamics simulations as a joint distribution of structural embeddings, capturing molecular movements between discrete time steps,” says Chalmers University of Technology associate professor Simon Olsson, who wasn’t involved in the research. “Leveraging a masked learning objective, MDGen enables innovative use cases such as transition path sampling, drawing analogies to inpainting trajectories connecting metastable phases.”

The researchers’ work on MDGen was supported, in part, by the National Institute of General Medical Sciences, the U.S. Department of Energy, the National Science Foundation, the Machine Learning for Pharmaceutical Discovery and Synthesis Consortium, the Abdul Latif Jameel Clinic for Machine Learning in Health, the Defense Threat Reduction Agency, and the Defense Advanced Research Projects Agency.

MIT physicists have created a new ultrathin, two-dimensional material with unusual magnetic properties that initially surprised the researchers before they went on to solve the complicated puzzle behind those properties’ emergence. As a result, the work introduces a new platform for studying how materials behave at the most fundamental level — the world of quantum physics.

Ultrathin materials made of a single layer of atoms have riveted scientists’ attention since the discovery of the first such material — graphene, composed of carbon — about 20 years ago. Among other advances since then, researchers have found that stacking individual sheets of the 2D materials, and sometimes twisting them at a slight angle to each other, can give them new properties, from superconductivity to magnetism. Enter the field of twistronics, which was pioneered at MIT by Pablo Jarillo-Herrero, the Cecil and Ida Green Professor of Physics at MIT.

In the current research, reported in the Jan. 7 issue of Nature Physics, the scientists, led by Jarillo-Herrero, worked with three layers of graphene. Each layer was twisted on top of the next at the same angle, creating a helical structure akin to the DNA helix or a hand of three cards that are fanned apart.

“Helicity is a fundamental concept in science, from basic physics to chemistry and molecular biology. With 2D materials, one can create special helical structures, with novel properties which we are just beginning to understand. This work represents a new twist in the field of twistronics, and the community is very excited to see what else we can discover using this helical materials platform!” says Jarillo-Herrero, who is also affiliated with MIT’s Materials Research Laboratory.

Do the twist

Twistronics can lead to new properties in ultrathin materials because arranging sheets of 2D materials in this way results in a unique pattern called a moiré lattice. And a moiré pattern, in turn, has an impact on the behavior of electrons.

“It changes the spectrum of energy levels available to the electrons and can provide the conditions for interesting phenomena to arise,” says Sergio C. de la Barrera, one of three co-first authors of the recent paper. De la Barrera, who conducted the work while a postdoc at MIT, is now an assistant professor at the University of Toronto.

In the current work, the helical structure created by the three graphene layers forms two moiré lattices. One is created by the first two overlapping sheets; the other is formed between the second and third sheets.

The two moiré patterns together form a third moiré, a supermoiré, or “moiré of a moiré,” says Li-Qiao Xia, a graduate student in MIT physics and another of the three co-first authors of the Nature Physics paper. “It’s like a moiré hierarchy.” While the first two moiré patterns are only nanometers, or billionths of a meter, in scale, the supermoiré appears at a scale of hundreds of nanometers superimposed over the other two. You can only see it if you zoom out to get a much wider view of the system.

A major surprise

The physicists expected to observe signatures of this moiré hierarchy. They got a huge surprise, however, when they applied and varied a magnetic field. The system responded with an experimental signature for magnetism, one that arises from the motion of electrons. In fact, this orbital magnetism persisted to -263 degrees Celsius — the highest temperature reported in carbon-based materials to date.

But that magnetism can only occur in a system that lacks a specific symmetry — one that the team’s new material should have had. “So the fact that we saw this was very puzzling. We didn’t really understand what was going on,” says Aviram Uri, an MIT Pappalardo postdoc in physics and the third co-first author of the new paper.

Other authors of the paper include MIT professor of physics Liang Fu; Aaron Sharpe of Sandia National Laboratories; Yves H. Kwan of Princeton University; Ziyan Zhu, David Goldhaber-Gordon, and Trithep Devakul of Stanford University; and Kenji Watanabe and Takashi Taniguchi of the National Institute for Materials Science in Japan.

What was happening?

It turns out that the new system did indeed break the symmetry that prohibits the orbital magnetism the team observed, but in a very unusual way. “What happens is that the atoms in this system aren’t very comfortable, so they move in a subtle orchestrated way that we call lattice relaxation,” says Xia. And the new structure formed by that relaxation does indeed break the symmetry locally, on the moiré length scale.

This opens the possibility for the orbital magnetism the team observed. However, if you zoom out to view the system on the supermoiré scale, the symmetry is restored. “The moiré hierarchy turns out to support interesting phenomena at different length scales,” says de la Barrera.

Concludes Uri: “It’s a lot of fun when you solve a riddle and it’s such an elegant solution. We’ve gained new insights into how electrons behave in these complex systems, insights that we couldn’t have had unless our experimental observations forced to think about these things.”

This work was supported by the Army Research Office, the National Science Foundation, the Gordon and Betty Moore Foundation, the Ross M. Brown Family Foundation, an MIT Pappalardo Fellowship, the VATAT Outstanding Postdoctoral Fellowship in Quantum Science and Technology, the JSPS KAKENHI, and a Stanford Science Fellowship. This work was carried out, in part, through the use of MIT.nano facilities.

A new experimental vaccine developed by researchers at MIT and Caltech could offer protection against emerging variants of SARS-CoV-2, as well as related coronaviruses, known as sarbecoviruses, that could spill over from animals to humans.

In addition to SARS-CoV-2, the virus that causes COVID-19, sarbecoviruses — a subgenus of coronaviruses — include the virus that led to the outbreak of the original SARS in the early 2000s. Sarbecoviruses that currently circulate in bats and other mammals may also hold the potential to spread to humans in the future.

By attaching up to eight different versions of sarbecovirus receptor-binding proteins (RBDs) to nanoparticles, the researchers created a vaccine that generates antibodies that recognize regions of RBDs that tend to remain unchanged across all strains of the viruses. That makes it much more difficult for viruses to evolve to escape vaccine-induced antibodies.

“This work is an example of how bringing together computation and immunological experiments can be fruitful,” says Arup K. 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.

Chakraborty and Pamela Bjorkman, a professor of biology and biological engineering at Caltech, are the senior authors of the study, which appears today in Cell. The paper’s lead authors are Eric Wang PhD ’24, Caltech postdoc Alexander Cohen, and Caltech graduate student Luis Caldera.

Mosaic nanoparticles

The new study builds on a project begun in Bjorkman’s lab, in which she and Cohen created a “mosaic” 60-mer nanoparticle that presents eight different sarbecovirus RBD proteins. The RBD is the part of the viral spike protein that helps the virus get into host cells. It is also the region of the coronavirus spike protein that is usually targeted by antibodies against sarbecoviruses.

RBDs contain some regions that are variable and can easily mutate to escape antibodies. Most of the antibodies generated by mRNA COVID-19 vaccines target those variable regions because they are more easily accessible. That is one reason why mRNA vaccines need to be updated to keep up with the emergence of new strains.

If researchers could create a vaccine that stimulates production of antibodies that target RBD regions that can’t easily change and are shared across viral strains, it could offer broader protection against a variety of sarbecoviruses.

Such a vaccine would have to stimulate B cells that have receptors (which then become antibodies) that target those shared, or “conserved,” regions. When B cells circulating in the body encounter a vaccine or other antigen, their B cell receptors, each of which have two “arms,” are more effectively activated if two copies of the antigen are available for binding to each arm. The conserved regions tend to be less accessible to B cell receptors, so if a nanoparticle vaccine presents just one type of RBD, B cells with receptors that bind to the more accessible variable regions, are most likely to be activated.

To overcome this, the Caltech researchers designed a nanoparticle vaccine that includes 60 copies of RBDs from eight different related sarbecoviruses, which have different variable regions but similar conserved regions. Because eight different RBDs are displayed on each nanoparticle, it’s unlikely that two identical RBDs will end up next to each other. Therefore, when a B cell receptor encounters the nanoparticle immunogen, the B cell is more likely to become activated if its receptor can recognize the conserved regions of the RBD.

“The concept behind the vaccine is that by co-displaying all these different RBDs on the nanoparticle, you are selecting for B cells that recognize the conserved regions that are shared between them,” Cohen says. “As a result, you’re selecting for B cells that are more cross-reactive. Therefore, the antibody response would be more cross-reactive and you could potentially get broader protection.”

In studies conducted in animals, the researchers showed that this vaccine, known as mosaic-8, produced strong antibody responses against diverse strains of SARS-CoV-2 and other sarbecoviruses and protected from challenges by both SARS-CoV-2 and SARS-CoV (original SARS).

Broadly neutralizing antibodies

After these studies were published in 2021 and 2022, the Caltech researchers teamed up with Chakraborty’s lab at MIT to pursue computational strategies that could allow them to identify RBD combinations that would generate even better antibody responses against a wider variety of sarbecoviruses.

Led by Wang, the MIT researchers pursued two different strategies — first, a large-scale computational screen of many possible mutations to the RBD of SARS-CoV-2, and second, an analysis of naturally occurring RBD proteins from zoonotic sarbecoviruses.

For the first approach, the researchers began with the original strain of SARS-CoV-2 and generated sequences of about 800,000 RBD candidates by making substitutions in locations that are known to affect antibody binding to variable portions of the RBD. Then, they screened those candidates for their stability and solubility, to make sure they could withstand attachment to the nanoparticle and injection as a vaccine.

From the remaining candidates, the researchers chose 10 based on how different their variable regions were. They then used these to create mosaic nanoparticles coated with either two or five different RBD proteins (mosaic-2_COM and mosaic-5_COM).

In their second approach, instead of mutating the RBD sequences, the researchers chose seven naturally occurring RBD proteins, using computational techniques to select RBDs that were different from each other in regions that are variable, but retained their conserved regions. They used these to create another vaccine, mosaic-7_COM.

Once the researchers produced the RBD-nanoparticles, they evaluated each one in mice. After each mouse received three doses of one of the vaccines, the researchers analyzed how well the resulting antibodies bound to and neutralized seven variants of SARS-CoV-2 and four other sarbecoviruses. 

They also compared the mosaic nanoparticle vaccines to a nanoparticle with only one type of RBD displayed, and to the original mosaic-8 particle from their 2021, 2022, and 2024 studies. They found that mosaic-2_COM and mosaic-5_COM outperformed both of those vaccines, and mosaic-7_COM showed the best responses of all. Mosaic-7_COM elicited antibodies with binding to most of the viruses tested, and these antibodies were also able to prevent the viruses from entering cells.

The researchers saw similar results when they tested the new vaccines in mice that were previously vaccinated with a bivalent mRNA COVID-19 vaccine.

“We wanted to simulate the fact that people have already been infected and/or vaccinated against SARS-CoV-2,” Wang says. “In pre-vaccinated mice, mosaic-7_COM is consistently giving the highest binding titers for both SARS-CoV-2 variants and other sarbecoviruses.”

Bjorkman’s lab has received funding from the Coalition for Epidemic Preparedness Innovations to do a clinical trial of the mosaic-8 RBD-nanoparticle. They also hope to move mosaic-7_COM, which performed better in the current study, into clinical trials. The researchers plan to work on redesigning the vaccines so that they could be delivered as mRNA, which would make them easier to manufacture.

The research was funded by a National Science Foundation Graduate Research Fellowship, the National Institutes of Health, Wellcome Leap, the Bill and Melinda Gates Foundation, the Coalition for Epidemic Preparedness Innovations, and the Caltech Merkin Institute for Translational Research.

Materials like car tires, human tissues, and spider webs are diverse in composition, but all contain networks of interconnected strands. A long-standing question about the durability of these materials asks: What is the energy required to fracture these diverse networks? A recently published paper by MIT researchers offers new insights.

“Our findings reveal a simple, general law that governs the fracture energy of networks across various materials and length scales,” says Xuanhe Zhao, the Uncas and Helen Whitaker Professor and professor of mechanical engineering and civil and environmental engineering at MIT. “This discovery has significant implications for the design of new materials, structures, and metamaterials, allowing for the creation of systems that are incredibly tough, soft, and stretchable.”

Despite an established understanding of the importance of failure resistance in design of such networks, no existing physical model effectively linked strand mechanics and connectivity to predict bulk fracture — until now. This new research reveals a universal scaling law that bridges length scales and makes it possible to predict the intrinsic fracture energy of diverse networks.

“This theory helps us predict how much energy it takes to break these networks by advancing a crack,” says graduate student Chase Hartquist, one of the paper’s lead authors. “It turns out that you can design tougher versions of these materials by making the strands longer, more stretchable, or resistant to higher forces before breaking.”

To validate their results, the team 3D-printed a giant, stretchable network, allowing them to demonstrate fracture properties in practice. They found that despite the differences in the networks, they all followed a simple and predictable rule. Beyond the changes to the strands themselves, a network can also be toughened by connecting the strands into larger loops.

“By adjusting these properties, car tires could last longer, tissues could better resist injury, and spider webs could become more durable,” says Hartquist.

Shu Wang, a postdoc in Zhao’s lab and fellow lead author of the paper, called the research findings “an extremely fulfilling moment ... it meant that the same rules could be applied to describe a wide variety of materials, making it easier to design the best material for a given situation.”

The researchers explain that this work represents progress in an exciting and emerging field called “architected materials,” where the structure within the material itself gives it unique properties. They say the discovery sheds light on how to make these materials even tougher, by focusing on designing the segments within the architecture stronger and more stretchable. The strategy is adaptable for materials across fields and can be applied to improve durability of soft robotic actuators, enhance the toughness of engineered tissues, or even create resilient lattices for aerospace technology.

Their open-access paper, “Scaling Law for Intrinsic Fracture Energy of Diverse Stretchable Networks,” is available now in Physical Review X, a leading journal in interdisciplinary physics.

According to the International Energy Agency, aviation accounts for about 2 percent of global carbon dioxide emissions, and aviation emissions are expected to double by mid-century as demand for domestic and international air travel rises. To sharply reduce emissions in alignment with the Paris Agreement’s long-term goal to keep global warming below 1.5 degrees Celsius, the International Air Transport Association (IATA) has set a goal to achieve net-zero carbon emissions by 2050. Which raises the question: Are there technologically feasible and economically viable strategies to reach that goal within the next 25 years?

To begin to address that question, a team of researchers at the MIT Center for Sustainability Science and Strategy (CS3) and the MIT Laboratory for Aviation and the Environment has spent the past year analyzing aviation decarbonization options in Latin America, where air travel is expected to more than triple by 2050 and thereby double today’s aviation-related emissions in the region.

Chief among those options is the development and deployment of sustainable aviation fuel. Currently produced from low- and zero-carbon sources (feedstock) including municipal waste and non-food crops, and requiring practically no alteration of aircraft systems or refueling infrastructure, sustainable aviation fuel (SAF) has the potential to perform just as well as petroleum-based jet fuel with as low as 20 percent of its carbon footprint.

Focused on Brazil, Chile, Colombia, Ecuador, Mexico and Peru, the researchers assessed SAF feedstock availability, the costs of corresponding SAF pathways, and how SAF deployment would likely impact fuel use, prices, emissions, and aviation demand in each country. They also explored how efficiency improvements and market-based mechanisms could help the region to reach decarbonization targets. The team’s findings appear in a CS3 Special Report.

SAF emissions, costs, and sources

Under an ambitious emissions mitigation scenario designed to cap global warming at 1.5 C and raise the rate of SAF use in Latin America to 65 percent by 2050, the researchers projected aviation emissions to be reduced by about 60 percent in 2050 compared to a scenario in which existing climate policies are not strengthened. To achieve net-zero emissions by 2050, other measures would be required, such as improvements in operational and air traffic efficiencies, airplane fleet renewal, alternative forms of propulsion, and carbon offsets and removals.

As of 2024, jet fuel prices in Latin America are around $0.70 per liter. Based on the current availability of feedstocks, the researchers projected SAF costs within the six countries studied to range from $1.11 to $2.86 per liter. They cautioned that increased fuel prices could affect operating costs of the aviation sector and overall aviation demand unless strategies to manage price increases are implemented.

Under the 1.5 C scenario, the total cumulative capital investments required to build new SAF producing plants between 2025 and 2050 were estimated at $204 billion for the six countries (ranging from $5 billion in Ecuador to $84 billion in Brazil). The researchers identified sugarcane- and corn-based ethanol-to-jet fuel, palm oil- and soybean-based hydro-processed esters and fatty acids as the most promising feedstock sources in the near term for SAF production in Latin America.

“Our findings show that SAF offers a significant decarbonization pathway, which must be combined with an economy-wide emissions mitigation policy that uses market-based mechanisms to offset the remaining emissions,” says Sergey Paltsev, lead author of the report, MIT CS3 deputy director, and senior research scientist at the MIT Energy Initiative.

Recommendations

The researchers concluded the report with recommendations for national policymakers and aviation industry leaders in Latin America.

They stressed that government policy and regulatory mechanisms will be needed to create sufficient conditions to attract SAF investments in the region and make SAF commercially viable as the aviation industry decarbonizes operations. Without appropriate policy frameworks, SAF requirements will affect the cost of air travel. For fuel producers, stable, long-term-oriented policies and regulations will be needed to create robust supply chains, build demand for establishing economies of scale, and develop innovative pathways for producing SAF.

Finally, the research team recommended a region-wide collaboration in designing SAF policies. A unified decarbonization strategy among all countries in the region will help ensure competitiveness, economies of scale, and achievement of long-term carbon emissions-reduction goals.

“Regional feedstock availability and costs make Latin America a potential major player in SAF production,” says Angelo Gurgel, a principal research scientist at MIT CS3 and co-author of the study. “SAF requirements, combined with government support mechanisms, will ensure sustainable decarbonization while enhancing the region’s connectivity and the ability of disadvantaged communities to access air transport.”

Financial support for this study was provided by LATAM Airlines and Airbus.

With a more efficient method for artificial pollination, farmers in the future could grow fruits and vegetables inside multilevel warehouses, boosting yields while mitigating some of agriculture’s harmful impacts on the environment.

To help make this idea a reality, MIT researchers are developing robotic insects that could someday swarm out of mechanical hives to rapidly perform precise pollination. However, even the best bug-sized robots are no match for natural pollinators like bees when it comes to endurance, speed, and maneuverability.

Now, inspired by the anatomy of these natural pollinators, the researchers have overhauled their design to produce tiny, aerial robots that are far more agile and durable than prior versions.

The new bots can hover for about 1,000 seconds, which is more than 100 times longer than previously demonstrated. The robotic insect, which weighs less than a paperclip, can fly significantly faster than similar bots while completing acrobatic maneuvers like double aerial flips.

The revamped robot is designed to boost flight precision and agility while minimizing the mechanical stress on its artificial wing flexures, which enables faster maneuvers, increased endurance, and a longer lifespan.

The new design also has enough free space that the robot could carry tiny batteries or sensors, which could enable it to fly on its own outside the lab.

“The amount of flight we demonstrated in this paper is probably longer than the entire amount of flight our field has been able to accumulate with these robotic insects. With the improved lifespan and precision of this robot, we are getting closer to some very exciting applications, like assisted pollination,” says Kevin Chen, an associate professor in the Department of Electrical Engineering and Computer Science (EECS), head of the Soft and Micro Robotics Laboratory within the Research Laboratory of Electronics (RLE), and the senior author of an open-access paper on the new design.

Chen is joined on the paper by co-lead authors Suhan Kim and Yi-Hsuan Hsiao, who are EECS graduate students; as well as EECS graduate student Zhijian Ren and summer visiting student Jiashu Huang. The research appears today in Science Robotics.

Boosting performance

Prior versions of the robotic insect were composed of four identical units, each with two wings, combined into a rectangular device about the size of a microcassette.

“But there is no insect that has eight wings. In our old design, the performance of each individual unit was always better than the assembled robot,” Chen says.

This performance drop was partly caused by the arrangement of the wings, which would blow air into each other when flapping, reducing the lift forces they could generate.

The new design chops the robot in half. Each of the four identical units now has one flapping wing pointing away from the robot’s center, stabilizing the wings and boosting their lift forces. With half as many wings, this design also frees up space so the robot could carry electronics.

In addition, the researchers created more complex transmissions that connect the wings to the actuators, or artificial muscles, that flap them. These durable transmissions, which required the design of longer wing hinges, reduce the mechanical strain that limited the endurance of past versions.

“Compared to the old robot, we can now generate control torque three times larger than before, which is why we can do very sophisticated and very accurate path-finding flights,” Chen says.

Yet even with these design innovations, there is still a gap between the best robotic insects and the real thing. For instance, a bee has only two wings, yet it can perform rapid and highly controlled motions.

“The wings of bees are finely controlled by a very sophisticated set of muscles. That level of fine-tuning is something that truly intrigues us, but we have not yet been able to replicate,” he says.

Less strain, more force

The motion of the robot’s wings is driven by artificial muscles. These tiny, soft actuators are made from layers of elastomer sandwiched between two very thin carbon nanotube electrodes and then rolled into a squishy cylinder. The actuators rapidly compress and elongate, generating mechanical force that flaps the wings.

In previous designs, when the actuator’s movements reach the extremely high frequencies needed for flight, the devices often start buckling. That reduces the power and efficiency of the robot. The new transmissions inhibit this bending-buckling motion, which reduces the strain on the artificial muscles and enables them to apply more force to flap the wings.

Another new design involves a long wing hinge that reduces torsional stress experienced during the flapping-wing motion. Fabricating the hinge, which is about 2 centimeters long but just 200 microns in diameter, was among their greatest challenges.

“If you have even a tiny alignment issue during the fabrication process, the wing hinge will be slanted instead of rectangular, which affects the wing kinematics,” Chen says.

After many attempts, the researchers perfected a multistep laser-cutting process that enabled them to precisely fabricate each wing hinge.

With all four units in place, the new robotic insect can hover for more than 1,000 seconds, which equates to almost 17 minutes, without showing any degradation of flight precision.

“When my student Nemo was performing that flight, he said it was the slowest 1,000 seconds he had spent in his entire life. The experiment was extremely nerve-racking,” Chen says.

The new robot also reached an average speed of 35 centimeters per second, the fastest flight researchers have reported, while performing body rolls and double flips. It can even precisely track a trajectory that spells M-I-T.

“At the end of the day, we’ve shown flight that is 100 times longer than anyone else in the field has been able to do, so this is an extremely exciting result,” he says.

From here, Chen and his students want to see how far they can push this new design, with the goal of achieving flight for longer than 10,000 seconds.

They also want to improve the precision of the robots so they could land and take off from the center of a flower. In the long run, the researchers hope to install tiny batteries and sensors onto the aerial robots so they could fly and navigate outside the lab.

“This new robot platform is a major result from our group and leads to many exciting directions. For example, incorporating sensors, batteries, and computing capabilities on this robot will be a central focus in the next three to five years,” Chen says.

This research is funded, in part, by the U.S. National Science Foundation and a Mathworks Fellowship.

Nearly 50 years ago, neuroscientists discovered cells within the brain’s hippocampus that store memories of specific locations. These cells also play an important role in storing memories of events, known as episodic memories. While the mechanism of how place cells encode spatial memory has been well-characterized, it has remained a puzzle how they encode episodic memories.

A new model developed by MIT researchers explains how those place cells can be recruited to form episodic memories, even when there’s no spatial component. According to this model, place cells, along with grid cells found in the entorhinal cortex, act as a scaffold that can be used to anchor memories as a linked series.

“This model is a first-draft model of the entorhinal-hippocampal episodic memory circuit. It’s a foundation to build on to understand the nature of episodic memory. That’s the thing I’m really excited about,” says Ila Fiete, a professor of brain and cognitive sciences at MIT, a member of MIT’s McGovern Institute for Brain Research, and the senior author of the new study.

The model accurately replicates several features of biological memory systems, including the large storage capacity, gradual degradation of older memories, and the ability of people who compete in memory competitions to store enormous amounts of information in “memory palaces.”

MIT Research Scientist Sarthak Chandra and Sugandha Sharma PhD ’24 are the lead authors of the study, which appears today in Nature. Rishidev Chaudhuri, an assistant professor at the University of California at Davis, is also an author of the paper.

An index of memories

To encode spatial memory, place cells in the hippocampus work closely with grid cells — a special type of neuron that fires at many different locations, arranged geometrically in a regular pattern of repeating triangles. Together, a population of grid cells forms a lattice of triangles representing a physical space.

In addition to helping us recall places where we’ve been, these hippocampal-entorhinal circuits also help us navigate new locations. From human patients, it’s known that these circuits are also critical for forming episodic memories, which might have a spatial component but mainly consist of events, such as how you celebrated your last birthday or what you had for lunch yesterday.

“The same hippocampal and entorhinal circuits are used not just for spatial memory, but also for general episodic memory,” Fiete says. “The question you can ask is what is the connection between spatial and episodic memory that makes them live in the same circuit?”

Two hypotheses have been proposed to account for this overlap in function. One is that the circuit is specialized to store spatial memories because those types of memories — remembering where food was located or where predators were seen — are important to survival. Under this hypothesis, this circuit encodes episodic memories as a byproduct of spatial memory.

An alternative hypothesis suggests that the circuit is specialized to store episodic memories, but also encodes spatial memory because location is one aspect of many episodic memories.

In this work, Fiete and her colleagues proposed a third option: that the peculiar tiling structure of grid cells and their interactions with hippocampus are equally important for both types of memory — episodic and spatial. To develop their new model, they built on computational models that her lab has been developing over the past decade, which mimic how grid cells encode spatial information.

“We reached the point where I felt like we understood on some level the mechanisms of the grid cell circuit, so it felt like the time to try to understand the interactions between the grid cells and the larger circuit that includes the hippocampus,” Fiete says.

In the new model, the researchers hypothesized that grid cells interacting with hippocampal cells can act as a scaffold for storing either spatial or episodic memory. Each activation pattern within the grid defines a “well,” and these wells are spaced out at regular intervals. The wells don’t store the content of a specific memory, but each one acts as a pointer to a specific memory, which is stored in the synapses between the hippocampus and the sensory cortex.

When the memory is triggered later from fragmentary pieces, grid and hippocampal cell interactions drive the circuit state into the nearest well, and the state at the bottom of the well connects to the appropriate part of the sensory cortex to fill in the details of the memory. The sensory cortex is much larger than the hippocampus and can store vast amounts of memory.

“Conceptually, we can think about the hippocampus as a pointer network. It’s like an index that can be pattern-completed from a partial input, and that index then points toward sensory cortex, where those inputs were experienced in the first place,” Fiete says. “The scaffold doesn’t contain the content, it only contains this index of abstract scaffold states.”

Furthermore, events that occur in sequence can be linked together: Each well in the grid cell-hippocampal network efficiently stores the information that is needed to activate the next well, allowing memories to be recalled in the right order.

Modeling memory cliffs and palaces

The researchers’ new model replicates several memory-related phenomena much more accurately than existing models that are based on Hopfield networks — a type of neural network that can store and recall patterns.

While Hopfield networks offer insight into how memories can be formed by strengthening connections between neurons, they don’t perfectly model how biological memory works. In Hopfield models, every memory is recalled in perfect detail until capacity is reached. At that point, no new memories can form, and worse, attempting to add more memories erases all prior ones. This “memory cliff” doesn’t accurately mimic what happens in the biological brain, which tends to gradually forget the details of older memories while new ones are continually added.

The new MIT model captures findings from decades of recordings of grid and hippocampal cells in rodents made as the animals explore and forage in various environments. It also helps to explain the underlying mechanisms for a memorization strategy known as a memory palace. One of the tasks in memory competitions is to memorize the shuffled sequence of cards in one or several card decks. They usually do this by assigning each card to a particular spot in a memory palace — a memory of a childhood home or other environment they know well. When they need to recall the cards, they mentally stroll through the house, visualizing each card in its spot as they go along. Counterintuitively, adding the memory burden of associating cards with locations makes recall stronger and more reliable.

The MIT team’s computational model was able to perform such tasks very well, suggesting that memory palaces take advantage of the memory circuit’s own strategy of associating inputs with a scaffold in the hippocampus, but one level down: Long-acquired memories reconstructed in the larger sensory cortex can now be pressed into service as a scaffold for new memories. This allows for the storage and recall of many more items in a sequence than would otherwise be possible.

The researchers now plan to build on their model to explore how episodic memories could become converted to cortical “semantic” memory, or the memory of facts dissociated from the specific context in which they were acquired (for example, Paris is the capital of France), how episodes are defined, and how brain-like memory models could be integrated into modern machine learning.

The research was funded by the U.S. Office of Naval Research, the National Science Foundation under the Robust Intelligence program, the ARO-MURI award, the Simons Foundation, and the K. Lisa Yang ICoN Center.

Quantum computing promises to solve complex problems exponentially faster than a classical computer, by using the principles of quantum mechanics to encode and manipulate information in quantum bits (qubits).

Qubits are the building blocks of a quantum computer. One challenge to scaling, however, is that qubits are highly sensitive to background noise and control imperfections, which introduce errors into the quantum operations and ultimately limit the complexity and duration of a quantum algorithm. To improve the situation, MIT researchers and researchers worldwide have continually focused on improving qubit performance. 

In new work, using a superconducting qubit called fluxonium, MIT researchers in the Department of Physics, the Research Laboratory of Electronics (RLE), and the Department of Electrical Engineering and Computer Science (EECS) developed two new control techniques to achieve a world-record single-qubit fidelity of 99.998 percent. This result complements then-MIT researcher Leon Ding’s demonstration last year of a 99.92 percent two-qubit gate fidelity. 

The paper’s senior authors are David Rower PhD ’24, a recent physics postdoc in MIT’s Engineering Quantum Systems (EQuS) group and now a research scientist at the Google Quantum AI laboratory; Leon Ding PhD ’23 from EQuS, now leading the Calibration team at Atlantic Quantum; and William D. Oliver, the Henry Ellis Warren Professor of EECS and professor of physics, leader of EQuS, director of the Center for Quantum Engineering, and RLE associate director. The paper recently appeared in the journal PRX Quantum.

Decoherence and counter-rotating errors

A major challenge with quantum computation is decoherence, a process by which qubits lose their quantum information. For platforms such as superconducting qubits, decoherence stands in the way of realizing higher-fidelity quantum gates.

Quantum computers need to achieve high gate fidelities in order to implement sustained computation through protocols like quantum error correction. The higher the gate fidelity, the easier it is to realize practical quantum computing.

MIT researchers are developing techniques to make quantum gates, the basic operations of a quantum computer, as fast as possible in order to reduce the impact of decoherence. However, as gates get faster, another type of error, arising from counter-rotating dynamics, can be introduced because of the way qubits are controlled using electromagnetic waves. 

Single-qubit gates are usually implemented with a resonant pulse, which induces Rabi oscillations between the qubit states. When the pulses are too fast, however, “Rabi gates” are not so consistent, due to unwanted errors from counter-rotating effects. The faster the gate, the more the counter-rotating error is manifest. For low-frequency qubits such as fluxonium, counter-rotating errors limit the fidelity of fast gates.

“Getting rid of these errors was a fun challenge for us,” says Rower. “Initially, Leon had the idea to utilize circularly polarized microwave drives, analogous to circularly polarized light, but realized by controlling the relative phase of charge and flux drives of a superconducting qubit. Such a circularly polarized drive would ideally be immune to counter-rotating errors.”

While Ding’s idea worked immediately, the fidelities achieved with circularly polarized drives were not as high as expected from coherence measurements.

“Eventually, we stumbled on a beautifully simple idea,” says Rower. “If we applied pulses at exactly the right times, we should be able to make counter-rotating errors consistent from pulse-to-pulse. This would make the counter-rotating errors correctable. Even better, they would be automatically accounted for with our usual Rabi gate calibrations!”

They called this idea “commensurate pulses,” since the pulses needed to be applied at times commensurate with intervals determined by the qubit frequency through its inverse, the time period. Commensurate pulses are defined simply by timing constraints and can be applied to a single linear qubit drive. In contrast, circularly polarized microwaves require two drives and some extra calibration.

“I had much fun developing the commensurate technique,” says Rower. “It was simple, we understood why it worked so well, and it should be portable to any qubit suffering from counter-rotating errors!”

“This project makes it clear that counter-rotating errors can be dealt with easily. This is a wonderful thing for low-frequency qubits such as fluxonium, which are looking more and more promising for quantum computing.”

Fluxonium’s promise

Fluxonium is a type of superconducting qubit made up of a capacitor and Josephson junction; unlike transmon qubits, however, fluxonium also includes a large “superinductor,” which by design helps protect the qubit from environmental noise. This results in performing logical operations, or gates, with greater accuracy.

Despite having higher coherence, however, fluxonium has a lower qubit frequency that is generally associated with proportionally longer gates.

“Here, we’ve demonstrated a gate that is among the fastest and highest-fidelity across all superconducting qubits,” says Ding. “Our experiments really show that fluxonium is a qubit that supports both interesting physical explorations and also absolutely delivers in terms of engineering performance.”

With further research, they hope to reveal new limitations and yield even faster and higher-fidelity gates.

“Counter-rotating dynamics have been understudied in the context of superconducting quantum computing because of how well the rotating-wave approximation holds in common scenarios,” says Ding. “Our paper shows how to precisely calibrate fast, low-frequency gates where the rotating-wave approximation does not hold.”

Physics and engineering team up

“This is a wonderful example of the type of work we like to do in EQuS, because it leverages fundamental concepts in both physics and electrical engineering to achieve a better outcome,” says Oliver. “It builds on our earlier work with non-adiabatic qubit control, applies it to a new qubit — fluxonium — and makes a beautiful connection with counter-rotating dynamics.”

The science and engineering teams enabled the high fidelity in two ways. First, the team demonstrated “commensurate” (synchronous) non-adiabatic control, which goes beyond the standard “rotating wave approximation” of standard Rabi approaches. This leverages ideas that won the 2023 Nobel Prize in Physics for ultrafast “attosecond” pulses of light.

Secondly, they demonstrated it using an analog to circularly polarized light. Rather than a physical electromagnetic field with a rotating polarization vector in real x-y space, they realized a synthetic version of circularly polarized light using the qubit’s x-y space, which in this case corresponds to its magnetic flux and electric charge.

The combination of a new take on an existing qubit design (fluxonium) and the application of advanced control methods applied to an understanding of the underlying physics enabled this result.

Platform-independent and requiring no additional calibration overhead, this work establishes straightforward strategies for mitigating counter-rotating effects from strong drives in circuit quantum electrodynamics and other platforms, which the researchers expect to be helpful in the effort to realize high-fidelity control for fault-tolerant quantum computing.

Adds Oliver, “With the recent announcement of Google’s Willow quantum chip that demonstrated quantum error correction beyond threshold for the first time, this is a timely result, as we have pushed performance even higher. Higher-performant qubits will lead to lower overhead requirements for implementing error correction.”  

Other researchers on the paper are RLE’s Helin Zhang, Max Hays, Patrick M. Harrington, Ilan T. Rosen, Simon Gustavsson, Kyle Serniak, Jeffrey A. Grover, and Junyoung An, who is also with EECS; and MIT Lincoln Laboratory’s Jeffrey M. Gertler, Thomas M. Hazard, Bethany M. Niedzielski, and Mollie E. Schwartz.

This research was funded, in part, by the U.S. Army Research Office, the U.S. Department of Energy Office of Science, National Quantum Information Science Research Centers, Co-design Center for Quantum Advantage, U.S. Air Force, the U.S. Office of the Director of National Intelligence, and the U.S. National Science Foundation.  

Back in the old days — the really old days — the task of designing materials was laborious. Investigators, over the course of 1,000-plus years, tried to make gold by combining things like lead, mercury, and sulfur, mixed in what they hoped would be just the right proportions. Even famous scientists like Tycho Brahe, Robert Boyle, and Isaac Newton tried their hands at the fruitless endeavor we call alchemy.

Materials science has, of course, come a long way. For the past 150 years, researchers have had the benefit of the periodic table of elements to draw upon, which tells them that different elements have different properties, and one can’t magically transform into another. Moreover, in the past decade or so, machine learning tools have considerably boosted our capacity to determine the structure and physical properties of various molecules and substances. New research by a group led by Ju Li — the Tokyo Electric Power Company Professor of Nuclear Engineering at MIT and professor of materials science and engineering — offers the promise of a major leap in capabilities that can facilitate materials design. The results of their investigation are reported in a December 2024 issue of Nature Computational Science.

At present, most of the machine-learning models that are used to characterize molecular systems are based on density functional theory (DFT), which offers a quantum mechanical approach to determining the total energy of a molecule or crystal by looking at the electron density distribution — which is, basically, the average number of electrons located in a unit volume around each given point in space near the molecule. (Walter Kohn, who co-invented this theory 60 years ago, received a Nobel Prize in Chemistry for it in 1998.) While the method has been very successful, it has some drawbacks, according to Li: “First, the accuracy is not uniformly great. And, second, it only tells you one thing: the lowest total energy of the molecular system.”

“Couples therapy” to the rescue

His team is now relying on a different computational chemistry technique, also derived from quantum mechanics, known as coupled-cluster theory, or CCSD(T). “This is the gold standard of quantum chemistry,” Li comments. The results of CCSD(T) calculations are much more accurate than what you get from DFT calculations, and they can be as trustworthy as those currently obtainable from experiments. The problem is that carrying out these calculations on a computer is very slow, he says, “and the scaling is bad: If you double the number of electrons in the system, the computations become 100 times more expensive.” For that reason, CCSD(T) calculations have normally been limited to molecules with a small number of atoms — on the order of about 10. Anything much beyond that would simply take too long.

That’s where machine learning comes in. CCSD(T) calculations are first performed on conventional computers, and the results are then used to train a neural network with a novel architecture specially devised by Li and his colleagues. After training, the neural network can perform these same calculations much faster by taking advantage of approximation techniques. What’s more, their neural network model can extract much more information about a molecule than just its energy. “In previous work, people have used multiple different models to assess different properties,” says Hao Tang, an MIT PhD student in materials science and engineering. “Here we use just one model to evaluate all of these properties, which is why we call it a ‘multi-task’ approach.”

The “Multi-task Electronic Hamiltonian network,” or MEHnet, sheds light on a number of electronic properties, such as the dipole and quadrupole moments, electronic polarizability, and the optical excitation gap — the amount of energy needed to take an electron from the ground state to the lowest excited state. “The excitation gap affects the optical properties of materials,” Tang explains, “because it determines the frequency of light that can be absorbed by a molecule.” Another advantage of their CCSD-trained model is that it can reveal properties of not only ground states, but also excited states. The model can also predict the infrared absorption spectrum of a molecule related to its vibrational properties, where the vibrations of atoms within a molecule are coupled to each other, leading to various collective behaviors.

The strength of their approach owes a lot to the network architecture. Drawing on the work of MIT Assistant Professor Tess Smidt, the team is utilizing a so-called E(3)-equivariant graph neural network, says Tang, “in which the nodes represent atoms and the edges that connect the nodes represent the bonds between atoms. We also use customized algorithms that incorporate physics principles — related to how people calculate molecular properties in quantum mechanics — directly into our model.”

Testing, 1, 2 3

When tested on its analysis of known hydrocarbon molecules, the model of Li et al. outperformed DFT counterparts and closely matched experimental results taken from the published literature.

Qiang Zhu — a materials discovery specialist at the University of North Carolina at Charlotte (who was not part of this study) — is impressed by what’s been accomplished so far. “Their method enables effective training with a small dataset, while achieving superior accuracy and computational efficiency compared to existing models,” he says. “This is exciting work that illustrates the powerful synergy between computational chemistry and deep learning, offering fresh ideas for developing more accurate and scalable electronic structure methods.”

The MIT-based group applied their model first to small, nonmetallic elements — hydrogen, carbon, nitrogen, oxygen, and fluorine, from which organic compounds can be made — and has since moved on to examining heavier elements: silicon, phosphorus, sulfur, chlorine, and even platinum. After being trained on small molecules, the model can be generalized to bigger and bigger molecules. “Previously, most calculations were limited to analyzing hundreds of atoms with DFT and just tens of atoms with CCSD(T) calculations,” Li says. “Now we’re talking about handling thousands of atoms and, eventually, perhaps tens of thousands.”

For now, the researchers are still evaluating known molecules, but the model can be used to characterize molecules that haven’t been seen before, as well as to predict the properties of hypothetical materials that consist of different kinds of molecules. “The idea is to use our theoretical tools to pick out promising candidates, which satisfy a particular set of criteria, before suggesting them to an experimentalist to check out,” Tang says.

It’s all about the apps

Looking ahead, Zhu is optimistic about the possible applications. “This approach holds the potential for high-throughput molecular screening,” he says. “That’s a task where achieving chemical accuracy can be essential for identifying novel molecules and materials with desirable properties.”

Once they demonstrate the ability to analyze large molecules with perhaps tens of thousands of atoms, Li says, “we should be able to invent new polymers or materials” that might be used in drug design or in semiconductor devices. The examination of heavier transition metal elements could lead to the advent of new materials for batteries — presently an area of acute need.

The future, as Li sees it, is wide open. “It’s no longer about just one area,” he says. “Our ambition, ultimately, is to cover the whole periodic table with CCSD(T)-level accuracy, but at lower computational cost than DFT. This should enable us to solve a wide range of problems in chemistry, biology, and materials science. It’s hard to know, at present, just how wide that range might be.”

This work was supported by the Honda Research Institute. Hao Tang acknowledges support from the Mathworks Engineering Fellowship. The calculations in this work were performed, in part, on the Matlantis high-speed universal atomistic simulator, the Texas Advanced Computing Center, the MIT SuperCloud, and the National Energy Research Scientific Computing.

When sound waves reach the inner ear, neurons there pick up the vibrations and alert the brain. Encoded in their signals is a wealth of information that enables us to follow conversations, recognize familiar voices, appreciate music, and quickly locate a ringing phone or crying baby.

Neurons send signals by emitting spikes — brief changes in voltage that propagate along nerve fibers, also known as action potentials. Remarkably, auditory neurons can fire hundreds of spikes per second, and time their spikes with exquisite precision to match the oscillations of incoming sound waves.

With powerful new models of human hearing, scientists at MIT’s McGovern Institute for Brain Research have determined that this precise timing is vital for some of the most important ways we make sense of auditory information, including recognizing voices and localizing sounds.

The open-access findings, reported Dec. 4 in the journal Nature Communications, show how machine learning can help neuroscientists understand how the brain uses auditory information in the real world. MIT professor and McGovern investigator Josh McDermott, who led the research, explains that his team’s models better-equip researchers to study the consequences of different types of hearing impairment and devise more effective interventions.

Science of sound

The nervous system’s auditory signals are timed so precisely, researchers have long suspected that timing is important to our perception of sound. Sound waves oscillate at rates that determine their pitch: Low-pitched sounds travel in slow waves, whereas high-pitched sound waves oscillate more frequently. The auditory nerve that relays information from sound-detecting hair cells in the ear to the brain generates electrical spikes that correspond to the frequency of these oscillations. “The action potentials in an auditory nerve get fired at very particular points in time relative to the peaks in the stimulus waveform,” explains McDermott, who is also associate head of the MIT Department of Brain and Cognitive Sciences.

This relationship, known as phase-locking, requires neurons to time their spikes with sub-millisecond precision. But scientists haven’t really known how informative these temporal patterns are to the brain. Beyond being scientifically intriguing, McDermott says, the question has important clinical implications: “If you want to design a prosthesis that provides electrical signals to the brain to reproduce the function of the ear, it’s arguably pretty important to know what kinds of information in the normal ear actually matter,” he says.

This has been difficult to study experimentally; animal models can’t offer much insight into how the human brain extracts structure in language or music, and the auditory nerve is inaccessible for study in humans. So McDermott and graduate student Mark Saddler PhD ’24 turned to artificial neural networks.

Artificial hearing

Neuroscientists have long used computational models to explore how sensory information might be decoded by the brain, but until recent advances in computing power and machine learning methods, these models were limited to simulating simple tasks. “One of the problems with these prior models is that they’re often way too good,” says Saddler, who is now at the Technical University of Denmark. For example, a computational model tasked with identifying the higher pitch in a pair of simple tones is likely to perform better than people who are asked to do the same thing. “This is not the kind of task that we do every day in hearing,” Saddler points out. “The brain is not optimized to solve this very artificial task.” This mismatch limited the insights that could be drawn from this prior generation of models.

To better understand the brain, Saddler and McDermott wanted to challenge a hearing model to do things that people use their hearing for in the real world, like recognizing words and voices. That meant developing an artificial neural network to simulate the parts of the brain that receive input from the ear. The network was given input from some 32,000 simulated sound-detecting sensory neurons and then optimized for various real-world tasks.

The researchers showed that their model replicated human hearing well — better than any previous model of auditory behavior, McDermott says. In one test, the artificial neural network was asked to recognize words and voices within dozens of types of background noise, from the hum of an airplane cabin to enthusiastic applause. Under every condition, the model performed very similarly to humans.

When the team degraded the timing of the spikes in the simulated ear, however, their model could no longer match humans’ ability to recognize voices or identify the locations of sounds. For example, while McDermott’s team had previously shown that people use pitch to help them identify people’s voices, the model revealed that that this ability is lost without precisely timed signals. “You need quite precise spike timing in order to both account for human behavior and to perform well on the task,” Saddler says. That suggests that the brain uses precisely timed auditory signals because they aid these practical aspects of hearing.

The team’s findings demonstrate how artificial neural networks can help neuroscientists understand how the information extracted by the ear influences our perception of the world, both when hearing is intact and when it is impaired. “The ability to link patterns of firing in the auditory nerve with behavior opens a lot of doors,” McDermott says.

“Now that we have these models that link neural responses in the ear to auditory behavior, we can ask, ‘If we simulate different types of hearing loss, what effect is that going to have on our auditory abilities?’” McDermott says. “That will help us better diagnose hearing loss, and we think there are also extensions of that to help us design better hearing aids or cochlear implants.” For example, he says, “The cochlear implant is limited in various ways — it can do some things and not others. What’s the best way to set up that cochlear implant to enable you to mediate behaviors? You can, in principle, use the models to tell you that.”

MIT physicists and colleagues have for the first time measured the geometry, or shape, of electrons in solids at the quantum level. Scientists have long known how to measure the energies and velocities of electrons in crystalline materials, but until now, those systems’ quantum geometry could only be inferred theoretically, or sometimes not at all.

The work, reported in the Nov. 25 issue of Nature Physics, “opens new avenues for understanding and manipulating the quantum properties of materials,” says Riccardo Comin, MIT’s Class of 1947 Career Development Associate Professor of Physics and leader of the work.

“We’ve essentially developed a blueprint for obtaining some completely new information that couldn’t be obtained before,” says Comin, who is also affiliated with MIT’s Materials Research Laboratory and the Research Laboratory of Electronics.

The work could be applied to “any kind of quantum material, not just the one we worked with,” says Mingu Kang PhD ’23, first author of the Nature Physics paper who conducted the work as an MIT graduate student and who is now a Kavli Postdoctoral Fellow at Cornell University’s Laboratory of Atomic and Solid State Physics. 

Kang was also invited to write an accompanying research briefing on the work, including its implications, for the Nov. 25 issue of Nature Physics.

A weird world

In the weird world of quantum physics, an electron can be described as both a point in space and a wave-like shape. At the heart of the current work is a fundamental object known as a wave function that describes the latter. “You can think of it like a surface in a three-dimensional space,” says Comin.

There are different types of wave functions, ranging from the simple to the complex. Think of a ball. That is analogous to a simple, or trivial, wave function. Now picture a Mobius strip, the kind of structure explored by M.C. Escher in his art. That’s analogous to a complex, or nontrivial, wave function. And the quantum world is filled with materials composed of the latter.

But until now, the quantum geometry of wave functions could only be inferred theoretically, or sometimes not at all. And the property is becoming more and more important as physicists find more and more quantum materials with potential applications in everything from quantum computers to advanced electronic and magnetic devices.

The MIT team solved the problem using a technique called angle-resolved photoemission spectroscopy, or ARPES. Comin, Kang, and some of the same colleagues had used the technique in other research. For example, in 2022 they reported discovering the “secret sauce” behind exotic properties of a new quantum material known as a kagome metal. That work, too, appeared in Nature Physics. In the current work, the team adapted ARPES to measure the quantum geometry of a kagome metal.

Close collaborations

Kang stresses that the new ability to measure the quantum geometry of materials “comes from the close cooperation between theorists and experimentalists.”

The Covid-19 pandemic, too, had an impact. Kang, who is from South Korea, was based in that country during the pandemic. “That facilitated a collaboration with theorists in South Korea,” says Kang, an experimentalist.

The pandemic also led to an unusual opportunity for Comin. He traveled to Italy to help run the ARPES experiments at the Italian Light Source Elettra, a national laboratory. The lab was closed during the pandemic, but was starting to reopen when Comin arrived. He found himself alone, however, when Kang tested positive for Covid and couldn’t join him. So he inadvertently ran the experiments himself with the support of local scientists. “As a professor, I lead projects, but students and postdocs actually carry out the work. So this is basically the last study where I actually contributed to the experiments themselves,” he says with a smile.

In addition to Kang and Comin, additional authors of the Nature Physics paper are Sunje Kim of Seoul National University (Kim is a co-first author with Kang); Paul M. Neves, a graduate student in the MIT Department of Physics; Linda Ye of Stanford University; Junseo Jung of Seoul National University; Denny Puntel of the University of Trieste; Federico Mazzola of Consiglio Nazionale delle Ricerche and Ca’ Foscari University of Venice; Shiang Fang of Google DeepMind; Chris Jozwiak, Aaron Bostwick, and Eli Rotenberg of Lawrence Berkeley National Laboratory; Jun Fuji and Ivana Vobornik of Consiglio Nazionale delle Ricerche; Jae-Hoon Park of Max Planck POSTECH/Korea Research Initiative and Pohang University of Science and Technology; Joseph G. Checkelsky, associate professor of physics at MIT; and Bohm-Jung Yang of Seoul National University, who co-led the research project with Comin.

This work was funded by the U.S. Air Force Office of Scientific Research, the U.S. National Science Foundation, the Gordon and Betty Moore Foundation, the National Research Foundation of Korea, the Samsung Science and Technology Foundation, the U.S. Army Research Office, the U.S. Department of Energy Office of Science, the Heising-Simons Physics Research Fellow Program, the Tsinghua Education Foundation, the NFFA-MUR Italy Progetti Internazionali facility, the Samsung Foundation of Culture, and the Kavli Institute at Cornell.

One supermassive black hole has kept astronomers glued to their scopes for the last several years. First came a surprise disappearance, and now, a precarious spinning act.

The black hole in question is 1ES 1927+654, which is about as massive as a million suns and sits in a galaxy that is 270 million light-years away. In 2018, astronomers at MIT and elsewhere observed that the black hole’s corona — a cloud of whirling, white-hot plasma — suddenly disappeared, before reassembling months later. The brief though dramatic shut-off was a first in black hole astronomy.

Members of the MIT team have now caught the same black hole exhibiting more unprecedented behavior.

The astronomers have detected flashes of X-rays coming from the black hole at a steadily increasing clip. Over a period of two years, the flashes, at millihertz frequencies, increased from every 18 minutes to every seven minutes. This dramatic speed-up in X-rays has not been seen from a black hole until now.

The researchers explored a number of scenarios for what might explain the flashes. They believe the most likely culprit is a spinning white dwarf — an extremely compact core of a dead star that is orbiting around the black hole and getting precariously closer to its event horizon, the boundary beyond which nothing can escape the black hole’s gravitational pull. If this is the case, the white dwarf must be pulling off an impressive balancing act, as it could be coming right up to the black hole’s edge without actually falling in.

“This would be the closest thing we know of around any black hole,” says Megan Masterson, a graduate student in physics at MIT, who co-led the discovery. “This tells us that objects like white dwarfs may be able to live very close to an event horizon for a relatively extended period of time.”

The researchers present their findings today at the 245th meeting of the American Astronomical Society.

If a white dwarf is at the root of the black hole’s mysterious flashing, it would also give off gravitational waves, in a range that would be detectable by next-generation observatories such as the European Space Agency's Laser Interferometer Space Antenna (LISA).

“These new detectors are designed to detect oscillations on the scale of minutes, so this black hole system is in that sweet spot,” says co-author Erin Kara, associate professor of physics at MIT.

The study’s other co-authors include MIT Kavli members Christos Panagiotou, Joheen Chakraborty, Kevin Burdge, Riccardo Arcodia, Ronald Remillard, and Jingyi Wang, along with collaborators from multiple other institutions.

Nothing normal

Kara and Masterson were part of the team that observed 1ES 1927+654 in 2018, as the black hole’s corona went dark, then slowly rebuilt itself over time. For a while, the newly reformed corona — a cloud of highly energetic plasma and X-rays — was the brightest X-ray-emitting object in the sky.

“It was still extremely bright, though it wasn’t doing anything new for a couple years and was kind of gurgling along. But we felt we had to keep monitoring it because it was so beautiful,” Kara says. “Then we noticed something that has never really been seen before.”

In 2022, the team looked through observations of the black hole taken by the European Space Agency’s XMM-Newton, a space-based observatory that detects and measures X-ray emissions from black holes, neutron stars, galactic clusters, and other extreme cosmic sources. They noticed that X-rays from the black hole appeared to pulse with increasing frequency. Such “quasi-periodic oscillations” have only been observed in a handful of other supermassive black holes, where X-ray flashes appear with regular frequency.

In the case of 1ES 1927+654, the flickering seemed to steadily ramp up, from every 18 minutes to every seven minutes over the span of two years.

“We’ve never seen this dramatic variability in the rate at which it’s flashing,” Masterson says. “This looked absolutely nothing like a normal supermassive black hole.”

The fact that the flashing was detected in the X-ray band points to the strong possibility that the source is somewhere very close to the black hole. The innermost regions of a black hole are extremely high-energy environments, where X-rays are produced by fast-moving, hot plasma. X-rays are less likely to be seen at farther distances, where gas can circle more slowly in an accretion disk. The cooler environment of the disk can emit optical and ultraviolet light, but rarely gives off X-rays.

“Seeing something in the X-rays is already telling you you’re pretty close to the black hole,” Kara says. “When you see variability on the timescale of minutes, that’s close to the event horizon, and the first thing your mind goes to is circular motion, and whether something could be orbiting around the black hole.”

X-ray kick-up

Whatever was producing the X-ray flashes was doing so at an extremely close distance from the black hole, which the researchers estimate to be within a few million miles of the event horizon.

Masterson and Kara explored models for various astrophysical phenomena that could explain the X-ray patterns that they observed, including a possibility relating to the black hole’s corona.

“One idea is that this corona is oscillating, maybe blobbing back and forth, and if it starts to shrink, those oscillations get faster as the scales get smaller,” Masterson says. “But we’re in the very early stages of understanding coronal oscillations.”

Another promising scenario, and one that scientists have a better grasp on in terms of the physics involved, has to do with a daredevil of a white dwarf. According to their modeling, the researchers estimate the white dwarf could have been about one-tenth the mass of the sun. In contrast, the supermassive black hole itself is on the order of 1 million solar masses.

When any object gets this close to a supermassive black hole, gravitational waves are expected to be emitted, dragging the object closer to the black hole. As it circles closer, the white dwarf moves at a faster rate, which can explain the increasing frequency of X-ray oscillations that the team observed.

The white dwarf is practically at the precipice of no return and is estimated to be just a few million miles from the event horizon. However, the researchers predict that the star will not fall in. While the black hole’s gravity may pull the white dwarf inward, the star is also shedding part of its outer layer into the black hole. This shedding acts as a small kick-back, such that the white dwarf — an incredibly compact object itself — can resist crossing the black hole’s boundary.

“Because white dwarfs are small and compact, they’re very difficult to shred apart, so they can be very close to a black hole,” Kara says. “If this scenario is correct, this white dwarf is right at the turn around point, and we may see it get further away.”

The team plans to continue observing the system, with existing and future telescopes, to better understand the extreme physics at work in a black hole’s innermost environments. They are particularly excited to study the system once the space-based gravitational-wave detector LISA launches — currently planned for the mid 2030s — as the gravitational waves that the system should give off will be in a sweet spot that LISA can clearly detect.

“The one thing I’ve learned with this source is to never stop looking at it because it will probably teach us something new,” Masterson says. “The next step is just to keep our eyes open.”

Many people around the globe are living in energy poverty, meaning they spend at least 8 percent of their annual household income on energy. Addressing this problem is not simple, but an experiment by MIT researchers shows that giving people better data about their energy use, plus some coaching on the subject, can lead them to substantially reduce their consumption and costs.

The experiment, based in Amsterdam, resulted in households cutting their energy expenses in half, on aggregate — a savings big enough to move three-quarters of them out of energy poverty.

“Our energy coaching project as a whole showed a 75 percent success rate at alleviating energy poverty,” says Joseph Llewellyn, a researcher with MIT’s Senseable City Lab and co-author of a newly published paper detailing the experiment’s results.

“Energy poverty afflicts families all over the world. With empirical evidence on which policies work, governments could focus their efforts more effectively,” says Fábio Duarte, associate director of MIT’s Senseable City Lab, and another co-author of the paper.

The paper, “Assessing the impact of energy coaching with smart technology interventions to alleviate energy poverty,” appears today in Nature Scientific Reports.

The authors are Llewellyn, who is also a researcher at the Amsterdam Institute for Advanced Metropolitan Solutions (AMS) and the KTH Royal Institute of Technology in Stockholm; Titus Venverloo, a research fellow at the MIT Senseable City Lab and AMS; Fábio Duarte, who is also a principal researcher MIT’s Senseable City Lab; Carlo Ratti, director of the Senseable City Lab; and Cecilia Katzeff, Fredrik Johansson, and Daniel Pargman of the KTH Royal Institute of Technology.

The researchers developed the study after engaging with city officials in Amsterdam. In the Netherlands, about 550,000 households, or 7 percent of the population, are considered to be in energy poverty; in the European Union, that figure is about 50 million. In the U.S., separate research has shown that about three in 10 households report trouble paying energy bills.

To conduct the experiment, the researchers ran two versions of an energy coaching intervention. In one version, 67 households received one report on their energy usage, along with coaching about how to increase energy efficiency. In the other version, 50 households received those things as well as a smart device giving them real-time updates on their energy consumption. (All households also received some modest energy-savings improvements at the outset, such as additional insulation.)

Across the two groups, homes typically reduced monthly consumption of electricity by 33 percent and gas by 42 percent. They lowered their bills by 53 percent, on aggregate, and the percentage of income they spent on energy dropped from 10.1 percent to 5.3 percent.

What were these households doing differently? Some of the biggest behavioral changes included things such as only heating rooms that were in use and unplugging devices not being used. Both of those changes save energy, but their benefits were not always understood by residents before they received energy coaching.

“The range of energy literacy was quite wide from one home to the next,” Llewellyn says. “And when I went somewhere as an energy coach, it was never to moralize about energy use. I never said, ‘Oh, you’re using way too much.’ It was always working on it with the households, depending on what people need for their homes.”

Intriguingly, the homes receiving the small devices that displayed real-time energy data only tended to use them for three or four weeks following a coaching visit. After that, people seemed to lose interest in very frequent monitoring of their energy use. And yet, a few weeks of consulting the devices tended to be long enough to get people to change their habits in a lasting way.

“Our research shows that smart devices need to be accompanied by a close understanding of what drives families to change their behaviors,” Venverloo says.

As the researchers acknowledge, working with consumers to reduce their energy consumption is just one way to help people escape energy poverty. Other “structural” factors that can help include lower energy prices and more energy-efficient buildings.

On the latter note, the current paper has given rise to a new experiment Llewellyn is developing with Amsterdam officials, to examine the benefits of retrofitting residental buildings to lower energy costs. In that case, local policymakers are trying to work out how to fund the retrofitting in such a way that landlords do not simply pass those costs on to tenants.

“We don’t want a household to save money on their energy bills if it also means the rent increases, because then we’ve just displaced expenses from one item to another,” Llewellyn says.

Households can also invest in products like better insulation themselves, for windows or heating components, although for low-income households, finding the money to pay for such things may not be trivial. That is especially the case, Llewellyn suggests, because energy costs can seem “invisible,” and a lower priority, than feeding and clothing a family.

“It’s a big upfront cost for a household that does not have 100 Euros to spend,” Llewellyn says. Compared to paying for other necessities, he notes, “Energy is often the thing that tends to fall last on their list. Energy is always going to be this invisible thing that hides behind the walls, and it’s not easy to change that.” 

On the first day of your vacation in a new city, your explorations expose you to innumerable individual places. While the memories of these spots (like a beautiful garden on a quiet side street) feel immediately indelible, it might be days before you have enough intuition about the neighborhood to direct a newer tourist to that same site and then maybe to the café you discovered nearby. A new study of mice by MIT neuroscientists at The Picower Insitute for Learning and Memory provides new evidence for how the brain forms cohesive cognitive maps of whole spaces and highlights the critical importance of sleep for the process.

Scientists have known for decades that the brain devotes neurons in a region called the hippocampus to remembering specific locations. So-called “place cells” reliably activate when an animal is at the location the neuron is tuned to remember. But more useful than having markers of specific spaces is having a mental model of how they all relate in a continuous overall geography. Though such “cognitive maps” were formally theorized in 1948, neuroscientists have remained unsure of how the brain constructs them. The new study in the December edition of Cell Reports finds that the capability may depend upon subtle but meaningful changes over days in the activity of cells that are only weakly attuned to individual locations, but that increase the robustness and refinement of the hippocampus’s encoding of the whole space. With sleep, the study’s analyses indicate, these “weakly spatial” cells increasingly enrich neural network activity in the hippocampus to link together these places into a cognitive map.

“On Day 1, the brain doesn’t represent the space very well,” says lead author Wei Guo, a research scientist in the lab of senior author Matthew Wilson, the Sherman Fairchild Professor in The Picower Institute and MIT’s departments of Biology and Brain and Cognitive Sciences. “Neurons represent individual locations, but together they don’t form a map. But on Day 5 they form a map. If you want a map, you need all these neurons to work together in a coordinated ensemble.”

Mice mapping mazes

To conduct the study, Guo and Wilson, along with labmates Jie “Jack” Zhang and Jonathan Newman, introduced mice to simple mazes of varying shapes and let them explore them freely for about 30 minutes a day for several days. Importantly, the mice were not directed to learn anything specific through the offer of any rewards. They just wandered. Previous studies have shown that mice naturally demonstrate “latent learning” of spaces from this kind of unrewarded experience after several days.

To understand how latent learning takes hold, Guo and his colleagues visually monitored hundreds of neurons in the CA1 area of the hippocampus by engineering cells to flash when a buildup of calcium ions made them electrically active. They not only recorded the neurons’ flashes when the mice were actively exploring, but also while they were sleeping. Wilson’s lab has shown that animals “replay” their previous journeys during sleep, essentially refining their memories by dreaming about their experiences.

Analysis of the recordings showed that the activity of the place cells developed immediately and remained strong and unchanged over several days of exploration. But this activity alone wouldn’t explain how latent learning or a cognitive map evolves over several days. So unlike in many other studies where scientists focus solely on the strong and clear activity of place cells, Guo extended his analysis to the more subtle and mysterious activity of cells that were not so strongly spatially tuned. 

Using an emerging technique called “manifold learning” he was able to discern that many of the “weakly spatial” cells gradually correlated their activity not with locations, but with activity patterns among other neurons in the network. As this was happening, Guo’s analyses showed, the network encoded a cognitive map of the maze that increasingly resembled the literal, physical space.

“Although not responding to specific locations like strongly spatial cells, weakly spatial cells specialize in responding to ‘‘mental locations,’’ i.e., specific ensemble firing patterns of other cells,” the study authors wrote. “If a weakly spatial cell’s mental field encompasses two subsets of strongly spatial cells that encode distinct locations, this weakly spatial cell can serve as a bridge between these locations.”

In other words, the activity of the weakly spatial cells likely stitches together the individual locations represented by the place cells into a mental map.

The need for sleep

Studies by Wilson’s lab and many others have shown that memories are consolidated, refined, and processed by neural activity, such as replay, that occurs during sleep and rest. Guo and Wilson’s team therefore sought to test whether sleep was necessary for the contribution of weakly spatial cells to latent learning of cognitive maps.

To do this they let some mice explore a new maze twice during the same day with a three-hour siesta in between. Some of the mice were allowed to sleep but some were not. The ones that did showed a significant refinement of their mental map, but the ones that weren’t allowed to sleep showed no such improvement. Not only did the network encoding of the map improve, but also measures of the tuning of individual cells during showed that sleep helped cells become better attuned both to places and to patterns of network activity, so-called “mental places” or “fields.”

Mental map meaning

The “cognitive maps” the mice encoded over several days were not literal, precise maps of the mazes, Guo notes. Instead they were more like schematics. Their value is that they provide the brain with a topology that can be explored mentally, without having to be in the physical space. For instance, once you’ve formed your cognitive map of the neighborhood around your hotel, you can plan the next morning’s excursion (e.g., you could imagine grabbing a croissant at the bakery you observed a few blocks west and then picture eating it on one of those benches you noticed in the park along the river).

Indeed, Wilson hypothesized that the weakly spatial cells’ activity may be overlaying salient non-spatial information that brings additional meaning to the maps (i.e., the idea of a bakery is not spatial, even if it’s closely linked to a specific location). The study, however, included no landmarks within the mazes and did not test any specific behaviors among the mice. But now that the study has identified that weakly spatial cells contribute meaningfully to mapping, Wilson said future studies can investigate what kind of information they may be incorporating into the animals’ sense of their environments. We seem to intuitively regard the spaces we inhabit as more than just sets of discrete locations.

“In this study we focused on animals behaving naturally and demonstrated that during freely exploratory behavior and subsequent sleep, in the absence of reinforcement, substantial neural plastic changes at the ensemble level still occur,” the authors concluded. “This form of implicit and unsupervised learning constitutes a crucial facet of human learning and intelligence, warranting further in-depth investigations.”

The Freedom Together Foundation, The Picower Institute, and the National Institutes of Health funded the study.

Does the United States have a “moral responsibility” for providing aid to poor nations — which have a significantly smaller carbon footprint and face catastrophic climate events at a much higher rate than wealthy countries?

A study published Dec. 11 in Climatic Change explores U.S. public opinion on global climate policies considering our nation’s historic role as a leading contributor of carbon emissions. The randomized, experimental survey specifically investigates American attitudes toward such a moral responsibility. 

The work was led by MIT Professor Evan Lieberman, the Total Chair on Contemporary African Politics and director of the MIT Center for International Studies, and Volha Charnysh, the Ford Career Development Associate Professor of Political Science, and was co-authored with MIT political science PhD student Jared Kalow and University of Pennsylvania postdoc Erin Walk PhD ’24. Here, Lieberman describes the team's research and insights, and offers recommendations that could result in more effective climate advocacy.

Q: What are the key findings — and any surprises — of your recent work on climate attitudes among the U.S. population?

A: A big question at the COP29 Climate talks in Baku, Azerbaijan was: Who will pay the trillions of dollars needed to help lower-income countries adapt to climate change? During past meetings, global leaders have come to an increasing consensus that the wealthiest countries should pay, but there has been little follow-through on commitments. In countries like the United States, popular opinion about such policies can weigh heavily on politicians' minds, as citizens focus on their own challenges at home.

Prime Minister Gaston Browne of Antigua and Barbuda is one of many who views such transfers as a matter of moral responsibility, explaining that many rich countries see climate finance as “a random act of charity ... not recognizing that they have a moral obligation to provide funding, especially the historical emitters and even those who currently have large emissions.”

In our study, we set out to measure American attitudes towards climate-related foreign aid, and explicitly to test the impact of this particular moral responsibility narrative. We did this on an experimental basis, so subjects were randomly assigned to receive different messages.

One message emphasized what we call a “climate justice” frame, and it argued that Americans should contribute to helping poor countries because of the United States’ disproportionate role in the emissions of greenhouse gasses that have led to global warming. That message had a positive impact on the extent to which citizens supported the use of foreign aid for climate adaptation in poor countries. However, when we looked at who was actually moved by the message, we found that the effect was larger and statistically significant only among Democrats, but not among Republicans.

We were surprised that a message emphasizing solidarity, the idea that “we are all in this together,” had no overall effect on citizen attitudes, Democrats or Republicans. 

Q: What are your recommendations toward addressing the attitudes on global climate policies within the U.S.?

A: First, given limited budgets and attention for communications campaigns, our research certainly suggests that emphasizing a bit of blaming and shaming is more powerful than more diffuse messages of shared responsibility.

But our research also emphasized how critically important it is to find new ways to communicate with Republicans about climate change and about foreign aid. Republicans were overwhelmingly less supportive of climate aid and yet even from that low baseline, a message that moved Democrats had a much more mixed reception among Republicans. Researchers and those working on the front lines of climate communications need to do more to better understand Republican perspectives. Younger Republicans, for example, might be more movable on key climate policies.

Q: With an incoming Trump administration, what are some of the specific hurdles and/or opportunities we face in garnering U.S. public support for international climate negotiations?

A: Not only did Trump demonstrate his disdain for international action on climate change by withdrawing from the Paris agreement during his first term in office, but he has indicated his intention to double down on such strategies in his second term. And the idea that he would support assistance for the world’s poorest countries harmed by climate change? This seems unlikely. Because we find Republican public opinion so firmly in line with these perspectives, frankly, it is hard to be optimistic.

Those Americans concerned with the effects of climate change may need to look to state-level, non-government, corporate, and more global organizations to support climate justice efforts.

Q: Are there any other takeaways you’d like to share?

A: Those working in the climate change area may need to rethink how we talk and message about the challenges the world faces. Right now, almost anything that sounds like “climate change” is likely to be rejected by Republican leaders and large segments of American society. Our approach of experimenting with different types of messages is a relatively low-cost strategy for identifying more promising strategies, targeted at Americans and at citizens in other wealthy countries.

But our study, in line with other work, also demonstrates that partisanship — identifying as a Republican or Democrat — is by far the strongest predictor of attitudes toward climate aid. While climate justice messaging can move attitudes slightly, the effects are still modest relative to the contributions of party identification itself. Just as Republican party elites were once persuaded to take leadership in the global fight against HIV and AIDS, a similar challenge lies ahead for climate aid.

Awed as a young child by the majesty of the Golden Gate Bridge in San Francisco, civil engineer and MIT Morningside Academy for Design (MAD) Fellow Zane Schemmer has retained his fascination with bridges: what they look like, why they work, and how they’re designed and built.

He weighed the choice between architecture and engineering when heading off to college, but, motivated by the why and how of structural engineering, selected the latter. Now he incorporates design as an iterative process in the writing of algorithms that perfectly balance the forces involved in discrete portions of a structure to create an overall design that optimizes function, minimizes carbon footprint, and still produces a manufacturable result.

While this may sound like an obvious goal in structural design, it’s not. It’s new. It’s a more holistic way of looking at the design process that can optimize even down to the materials, angles, and number of elements in the nodes or joints that connect the larger components of a building, bridge, tower, etc.

According to Schemmer, there hasn’t been much progress on optimizing structural design to minimize embodied carbon, and the work that exists often results in designs that are “too complex to be built in real life,” he says. The embodied carbon of a structure is the total carbon dioxide emissions of its life cycle: from the extraction or manufacture of its materials to their transport and use and through the demolition of the structure and disposal of the materials. Schemmer, who works with Josephine V. Carstensen, the Gilbert W. Winslow Career Development Associate Professor of Civil and Environmental Engineering at MIT, is focusing on the portion of that cycle that runs through construction.

In September, at the IASS 2024 symposium "Redefining the Art of Structural Design in Zurich," Schemmer and Carstensen presented their work on Discrete Topology Optimization algorithms that are able to minimize the embodied carbon in a bridge or other structure by up to 20 percent. This comes through materials selection that considers not only a material’s appearance and its ability to get the job done, but also the ease of procurement, its proximity to the building site, and the carbon embodied in its manufacture and transport.

“The real novelty of our algorithm is its ability to consider multiple materials in a highly constrained solution space to produce manufacturable designs with a user-specified force flow,” Schemmer says. “Real-life problems are complex and often have many constraints associated with them. In traditional formulations, it can be difficult to have a long list of complicated constraints. Our goal is to incorporate these constraints to make it easier to take our designs out of the computer and create them in real life.”

Take, for instance, a steel tower, which could be a “super lightweight, efficient design solution,” Schemmer explains. Because steel is so strong, you don’t need as much of it compared to concrete or timber to build a big building. But steel is also very carbon-intensive to produce and transport. Shipping it across the country or especially from a different continent can sharply increase its embodied carbon price tag. Schemmer’s topology optimization will replace some of the steel with timber elements or decrease the amount of steel in other elements to create a hybrid structure that will function effectively and minimize the carbon footprint. “This is why using the same steel in two different parts of the world can lead to two different optimized designs,” he explains.

Schemmer, who grew up in the mountains of Utah, earned a BS and MS in civil and environmental engineering from University of California at Berkeley, where his graduate work focused on seismic design. He describes that education as providing a “very traditional, super-strong engineering background that tackled some of the toughest engineering problems,” along with knowledge of structural engineering’s traditions and current methods.

But at MIT, he says, a lot of the work he sees “looks at removing the constraints of current societal conventions of doing things, and asks how could we do things if it was in a more ideal form; what are we looking at then? Which I think is really cool,” he says. “But I think sometimes too, there’s a jump between the most-perfect version of something and where we are now, that there needs to be a bridge between those two. And I feel like my education helps me see that bridge.”

The bridge he’s referring to is the topology optimization algorithms that make good designs better in terms of decreased global warming potential.

“That’s where the optimization algorithm comes in,” Schemmer says. “In contrast to a standard structure designed in the past, the algorithm can take the same design space and come up with a much more efficient material usage that still meets all the structural requirements, be up to code, and have everything we want from a safety standpoint.”

That’s also where the MAD Design Fellowship comes in. The program provides yearlong fellowships with full financial support to graduate students from all across the Institute who network with each other, with the MAD faculty, and with outside speakers who use design in new ways in a surprising variety of fields. This helps the fellows gain a better understanding of how to use iterative design in their own work.

“Usually people think of their own work like, ‘Oh, I had this background. I’ve been looking at this one way for a very long time.’ And when you look at it from an outside perspective, I think it opens your mind to be like, ‘Oh my God. I never would have thought about doing this that way. Maybe I should try that.’ And then we can move to new ideas, new inspiration for better work,” Schemmer says.

He chose civil and structural engineering over architecture some seven years ago, but says that “100 years ago, I don’t think architecture and structural engineering were two separate professions. I think there was an understanding of how things looked and how things worked, and it was merged together. Maybe from an efficiency standpoint, it’s better to have things done separately. But I think there’s something to be said for having knowledge about how the whole system works, potentially more intermingling between the free-form architectural design and the mathematical design of a civil engineer. Merging it back together, I think, has a lot of benefits.”

Which brings us back to the Golden Gate Bridge, Schemmer’s longtime favorite. You can still hear that excited 3-year-old in his voice when he talks about it.

“It’s so iconic,” he says. “It’s connecting these two spits of land that just rise straight up out of the ocean. There’s this fog that comes in and out a lot of days. It's a really magical place, from the size of the cable strands and everything. It’s just, ‘Wow.’ People built this over 100 years ago, before the existence of a lot of the computational tools that we have now. So, all the math, everything in the design, was all done by hand and from the mind. Nothing was computerized, which I think is crazy to think about.”

As Schemmer continues work on his doctoral degree at MIT, the MAD fellowship will expose him to many more awe-inspiring ideas in other fields, leading him to incorporate some of these in some way with his engineering knowledge to design better ways of building bridges and other structures.

Whether you’re describing the sound of your faulty car engine or meowing like your neighbor’s cat, imitating sounds with your voice can be a helpful way to relay a concept when words don’t do the trick.

Vocal imitation is the sonic equivalent of doodling a quick picture to communicate something you saw — except that instead of using a pencil to illustrate an image, you use your vocal tract to express a sound. This might seem difficult, but it’s something we all do intuitively: To experience it for yourself, try using your voice to mirror the sound of an ambulance siren, a crow, or a bell being struck.

Inspired by the cognitive science of how we communicate, MIT Computer Science and Artificial Intelligence Laboratory (CSAIL) researchers have developed an AI system that can produce human-like vocal imitations with no training, and without ever having "heard" a human vocal impression before.

To achieve this, the researchers engineered their system to produce and interpret sounds much like we do. They started by building a model of the human vocal tract that simulates how vibrations from the voice box are shaped by the throat, tongue, and lips. Then, they used a cognitively-inspired AI algorithm to control this vocal tract model and make it produce imitations, taking into consideration the context-specific ways that humans choose to communicate sound.

The model can effectively take many sounds from the world and generate a human-like imitation of them — including noises like leaves rustling, a snake’s hiss, and an approaching ambulance siren. Their model can also be run in reverse to guess real-world sounds from human vocal imitations, similar to how some computer vision systems can retrieve high-quality images based on sketches. For instance, the model can correctly distinguish the sound of a human imitating a cat’s “meow” versus its “hiss.”

In the future, this model could potentially lead to more intuitive “imitation-based” interfaces for sound designers, more human-like AI characters in virtual reality, and even methods to help students learn new languages.

The co-lead authors — MIT CSAIL PhD students Kartik Chandra SM ’23 and Karima Ma, and undergraduate researcher Matthew Caren — note that computer graphics researchers have long recognized that realism is rarely the ultimate goal of visual expression. For example, an abstract painting or a child’s crayon doodle can be just as expressive as a photograph.

“Over the past few decades, advances in sketching algorithms have led to new tools for artists, advances in AI and computer vision, and even a deeper understanding of human cognition,” notes Chandra. “In the same way that a sketch is an abstract, non-photorealistic representation of an image, our method captures the abstract, non-phono-realistic ways humans express the sounds they hear. This teaches us about the process of auditory abstraction.”

The art of imitation, in three parts

The team developed three increasingly nuanced versions of the model to compare to human vocal imitations. First, they created a baseline model that simply aimed to generate imitations that were as similar to real-world sounds as possible — but this model didn’t match human behavior very well.

The researchers then designed a second “communicative” model. According to Caren, this model considers what’s distinctive about a sound to a listener. For instance, you’d likely imitate the sound of a motorboat by mimicking the rumble of its engine, since that’s its most distinctive auditory feature, even if it’s not the loudest aspect of the sound (compared to, say, the water splashing). This second model created imitations that were better than the baseline, but the team wanted to improve it even more.

To take their method a step further, the researchers added a final layer of reasoning to the model. “Vocal imitations can sound different based on the amount of effort you put into them. It costs time and energy to produce sounds that are perfectly accurate,” says Chandra. The researchers’ full model accounts for this by trying to avoid utterances that are very rapid, loud, or high- or low-pitched, which people are less likely to use in a conversation. The result: more human-like imitations that closely match many of the decisions that humans make when imitating the same sounds.

After building this model, the team conducted a behavioral experiment to see whether the AI- or human-generated vocal imitations were perceived as better by human judges. Notably, participants in the experiment favored the AI model 25 percent of the time in general, and as much as 75 percent for an imitation of a motorboat and 50 percent for an imitation of a gunshot.

Toward more expressive sound technology

Passionate about technology for music and art, Caren envisions that this model could help artists better communicate sounds to computational systems and assist filmmakers and other content creators with generating AI sounds that are more nuanced to a specific context. It could also enable a musician to rapidly search a sound database by imitating a noise that is difficult to describe in, say, a text prompt.

In the meantime, Caren, Chandra, and Ma are looking at the implications of their model in other domains, including the development of language, how infants learn to talk, and even imitation behaviors in birds like parrots and songbirds.

The team still has work to do with the current iteration of their model: It struggles with some consonants, like “z,” which led to inaccurate impressions of some sounds, like bees buzzing. They also can’t yet replicate how humans imitate speech, music, or sounds that are imitated differently across different languages, like a heartbeat.

Stanford University linguistics professor Robert Hawkins says that language is full of onomatopoeia and words that mimic but don’t fully replicate the things they describe, like the “meow” sound that very inexactly approximates the sound that cats make. “The processes that get us from the sound of a real cat to a word like ‘meow’ reveal a lot about the intricate interplay between physiology, social reasoning, and communication in the evolution of language,” says Hawkins, who wasn’t involved in the CSAIL research. “This model presents an exciting step toward formalizing and testing theories of those processes, demonstrating that both physical constraints from the human vocal tract and social pressures from communication are needed to explain the distribution of vocal imitations.”

Caren, Chandra, and Ma wrote the paper with two other CSAIL affiliates: Jonathan Ragan-Kelley, MIT Department of Electrical Engineering and Computer Science associate professor, and Joshua Tenenbaum, MIT Brain and Cognitive Sciences professor and Center for Brains, Minds, and Machines member. Their work was supported, in part, by the Hertz Foundation and the National Science Foundation. It was presented at SIGGRAPH Asia in early December.

Researchers in MIT Professor Stefanie Mueller’s group have spent much of the last decade developing a variety of computing techniques aimed at reimagining how products and systems are designed. Much in the way that platforms like Instagram allow users to modify 2-D photographs with filters, Mueller imagines a world where we can do the same thing for a wide array of physical objects.

In a new open-access paper, her team at MIT’s Computer Science and Artificial Intelligence Laboratory (CSAIL) has demonstrated a novel printing technique along these lines — which they call “Thermochromorph” — that produces images that can change colors when heated up.

Led by first author and MIT electrical engineering and computer science doctoral student Ticha Melody Sethapakdi SM '22, the researchers say that they could imagine their method being applied in ways that are both artistic and functional, like a coffee-cup that warns if the liquid is too hot, or packaging for medicines or perishable foods that could indicate if the product has been stored at a safe temperature.

So-called “thermochromic” materials that visually change with temperature are not new — you can see examples with consumer beverages like Coke and Coors Light that reveal “ready to drink” labeling when refrigerated. But such instances in product marketing have traditionally been limited to a single color. By using inks with complementary characteristics — with one set that goes from clear to colored, and another from colored to clear — Sethapakdi says that she and her colleagues are “finally taking advantage of full-color process printing, which opens up a lot of possibilities for designing with thermochromic materials.”

The researchers worked with several visual artists to teach them to use Thermochromorph, and then solicited feedback and brainstorming about new narrative concepts and techniques unlocked by the tool, like color-changing postcards that could tell sequential stories in more compact, dynamic ways. One participant even plans to use Thermochromorph to make an educational science kit aimed at teaching students about sea creatures that change color.

The team developed their method to be applied specifically to “relief printing,” an early form of printmaking that involves carving a design into a block of material, applying ink or pigment to it, and then transferring the image onto paper or another surface.

Sethapakdi says that, compared to techniques like screen printing, relief printing is “more lightweight” and can be done with less setup and fewer materials, enabling a faster, lower-stakes iteration process. Artists that include the likes of Pablo Picasso and Salvador Dalí have used a range of related approaches in their work, such as woodcut and linocut printing.

“Our key contribution is applying these new materials to a traditional artistic process, and exploring how artists might be able to use it as part of their practice,” says Sethapakdi, lead author on a related paper that was recently presented at SIGGRAPH Asia in Tokyo.

The color-changing component also need not come from an active external heating or cooling source like, say, a fridge or a hot plate; using thermochromic inks with lower activation temperatures can allow for more subtle thermal changes brought about by human touch. Sethapakdi says she could even imagine applying this new process to create interactive surfaces or dynamic analog “interfaces” that visually change in response to touch.

Thermochromorph combines digital and analog processes in the form of, on the one hand, CMYK imaging and laser cutting, and, on the other, manual printmaking and thermochromic inks. Fabrication involves four core steps:
 

1. Block preparation: Solid hardwood blocks are used for Thermochromorph. The blocks are laser cut and engraved with the desired design, and then rinsed with water to remove any leftover particles.
2. Inking the block: First, a thin layer of ink is spread evenly onto a plate using a rubber brayer. Then, the ink is transferred from the brayer to the woodblock.
3. Registration: A registration jig is used to position the woodblock to ensure the different ink layers are aligned correctly. The printing surface, such as paper, is then placed on top of the block and secured.
4. Printing the images: A printing press is used to apply even pressure across the printing surface and transfer the ink from the block to the surface. The hot image is printed first, followed by the cold image. (If necessary, additional ink can be applied to specific areas of the block to touch up the print.)

The three prints the team used to demonstrate their technique were a set of frames from a Batman comic, a label depicting a fish and its underlying skeleton, and an image of a male subject both in profile and viewed from the front. (For the latter, as the temperature changes, the viewpoint gradually shifts, giving the effect of motion.)

It’s worth noting that Thermochromorph does have some potential limitations related to image resolution and print quality. Specifically, image resolution is constrained by the smallest dot size that the team’s laser cutter can engrave. Techniques like screen printing would offset this, but with the additional drawback of needing more time and materials. In terms of print quality, the pigments are not entirely invisible in their ‘clear’ states, which means that the clarity of the transitions depends on how thickly the ink layers were applied during printmaking. While this issue is intrinsic to the properties of the pigments, Sethapakdi says that for future iterations the team plans to explore different image-processing techniques to modify the overlay of halftone patterns for the hot and cold images, which may help to reduce these visual artifacts.

Sethapakdi and Mueller co-authored the new paper alongside Juliana Covarrubias ’24, MIT graduate student in media arts and sciences Paris Myers, University of California at Berkeley PhD student Tianyu Yu, and Adobe Research Scientist Mackenzie Leake.

A recent study from the McGovern Institute for Brain Research shows how interests can modulate language processing in children’s brains and paves the way for personalized brain research.

The paper, which appears in Imaging Neuroscience, was conducted in the lab of MIT professor and McGovern Institute investigator John Gabrieli, and led by senior author Anila D’Mello, a recent McGovern postdoc who is now an assistant professor at the University of Texas Southwestern Medical Center and the University of Texas at Dallas.

“Traditional studies give subjects identical stimuli to avoid confounding the results,” says Gabrieli, who is the Grover Hermann Professor of Health Sciences and Technology and a professor of brain and cognitive sciences at MIT. “However, our research tailored stimuli to each child’s interest, eliciting stronger — and more consistent — activity patterns in the brain’s language regions across individuals.” 

Funded by the Hock E. Tan and K. Lisa Yang Center for Autism Research in MIT’s Yang Tan Collective, this work unveils a new paradigm that challenges current methods and shows how personalization can be a powerful strategy in neuroscience. The paper’s co-first authors are Halie Olson, a postdoc at the McGovern Institute, and Kristina Johnson PhD '21, an assistant professor at Northeastern University and former doctoral student at the MIT Media Lab. “Our research integrates participants’ lived experiences into the study design,” says Johnson. “This approach not only enhances the validity of our findings, but also captures the diversity of individual perspectives, often overlooked in traditional research.”

Taking interest into account

When it comes to language, our interests are like operators behind the switchboard. They guide what we talk about and who we talk to. Research suggests that interests are also potent motivators and can help improve language skills. For instance, children score higher on reading tests when the material covers topics that are interesting to them.

But neuroscience has shied away from using personal interests to study the brain, especially in the realm of language. This is mainly because interests, which vary between people, could throw a wrench into experimental control — a core principle that drives scientists to limit factors that can muddle the results.

Gabrieli, D’Mello, Olson, and Johnson ventured into this unexplored territory. The team wondered if tailoring language stimuli to children’s interests might lead to higher responses in language regions of the brain. “Our study is unique in its approach to control the kind of brain activity our experiments yield, rather than control the stimuli we give subjects,” says D’Mello. “This stands in stark contrast to most neuroimaging studies that control the stimuli but might introduce differences in each subject’s level of interest in the material.”

In their recent study, the authors recruited a cohort of 20 children to investigate how personal interests affected the way the brain processes language. Caregivers described their child’s interests to the researchers, spanning baseball, train lines, “Minecraft,” and musicals. During the study, children listened to audio stories tuned to their unique interests. They were also presented with audio stories about nature (this was not an interest among the children) for comparison. To capture brain activity patterns, the team used functional magnetic resonance imaging (fMRI), which measures changes in blood flow caused by underlying neural activity.

New insights into the brain

“We found that, when children listened to stories about topics they were really interested in, they showed stronger neural responses in language areas than when they listened to generic stories that weren’t tailored to their interests,” says Olson. “Not only does this tell us how interests affect the brain, but it also shows that personalizing our experimental stimuli can have a profound impact on neuroimaging results.”

The researchers noticed a particularly striking result. “Even though the children listened to completely different stories, their brain activation patterns were more overlapping with their peers when they listened to idiosyncratic stories compared to when they listened to the same generic stories about nature,” says D’Mello. This, she notes, points to how interests can boost both the magnitude and consistency of signals in language regions across subjects without changing how these areas communicate with each other.

Gabrieli noted another finding: “In addition to the stronger engagement of language regions for content of interest, there was also stronger activation in brain regions associated with reward and also with self-reflection.” Personal interests are individually relevant and can be rewarding, potentially driving higher activation in these regions during personalized stories.

These personalized paradigms might be particularly well-suited to studies of the brain in unique or neurodivergent populations. Indeed, the team is already applying these methods to study language in the brains of autistic children.

This study breaks new ground in neuroscience and serves as a prototype for future work that personalizes research to unearth further knowledge of the brain. In doing so, scientists can compile a more complete understanding of the type of information that is processed by specific brain circuits and more fully grasp complex functions such as language. 

Researchers at MIT’s Laboratory for Information and Decision Systems (LIDS) have shown that using decision-making software and dynamic monitoring of weather and energy use can significantly improve resiliency in the face of weather-related outages, and can also help to efficiently integrate renewable energy sources into the grid.

The researchers point out that the system they suggest might have prevented or at least lessened the kind of widespread power outage that Puerto Rico experienced last week by providing analysis to guide rerouting of power through different lines and thus limit the spread of the outage.

The computer platform, which the researchers describe as DyMonDS, for Dynamic Monitoring and Decision Systems, can be used to enhance the existing operating and planning practices used in the electric industry. The platform supports interactive information exchange and decision-making between the grid operators and grid-edge users — all the distributed power sources, storage systems and software that contribute to the grid. It also supports optimization of available resources and controllable grid equipment as system conditions vary. It further lends itself to implementing cooperative decision-making by different utility- and non-utility-owned electric power grid users, including portfolios of mixed resources, users, and storage. Operating and planning the interactions of the end-to-end high-voltage transmission grid with local distribution grids and microgrids represents another major potential use of this platform.

This general approach was illustrated using a set of publicly-available data on both meteorology and details of electricity production and distribution in Puerto Rico. An extended AC Optimal Power Flow software developed by SmartGridz Inc. is used for system-level optimization of controllable equipment. This provides real-time guidance for deciding how much power, and through which transmission lines, should be channeled by adjusting plant dispatch and voltage-related set points, and in extreme cases, where to reduce or cut power in order to maintain physically-implementable service for as many customers as possible. The team found that the use of such a system can help to ensure that the greatest number of critical services maintain power even during a hurricane, and at the same time can lead to a substantial decrease in the need for construction of new power plants thanks to more efficient use of existing resources.

The findings are described in a paper in the journal Foundations and Trends in Electric Energy Systems, by MIT LIDS researchers Marija Ilic and Laurentiu Anton, along with recent alumna Ramapathi Jaddivada.

“Using this software,” Ilic says, they show that “even during bad weather, if you predict equipment failures, and by using that information exchange, you can localize the effect of equipment failures and still serve a lot of customers, 50 percent of customers, when otherwise things would black out.”

Anton says that “the way many grids today are operated is sub-optimal.” As a result, “we showed how much better they could do even under normal conditions, without any failures, by utilizing this software.” The savings resulting from this optimization, under everyday conditions, could be in the tens of percents, they say.

The way utility systems plan currently, Ilic says, “usually the standard is that they have to build enough capacity and operate in real time so that if one large piece of equipment fails, like a large generator or transmission line, you still serve customers in an uninterrupted way. That’s what’s called N-minus-1.” Under this policy, if one major component of the system fails, they should be able to maintain service for at least 30 minutes. That system allows utilities to plan for how much reserve generating capacity they need to have on hand. That’s expensive, Ilic points out, because it means maintaining this reserve capacity all the time, even under normal operating conditions when it’s not needed.

In addition, “right now there are no criteria for what I call N-minus-K,” she says. If bad weather causes five pieces of equipment to fail at once, “there is no software to help utilities decide what to schedule” in terms of keeping the most customers, and the most important services such as hospitals and emergency services, provided with power. They showed that even with 50 percent of the infrastructure out of commission, it would still be possible to keep power flowing to a large proportion of customers.

Their work on analyzing the power situation in Puerto Rico started after the island had been devastated by hurricanes Irma and Maria. Most of the electric generation capacity is in the south, yet the largest loads are in San Juan, in the north, and Mayaguez in the west. When transmission lines get knocked down, a lot of rerouting of power needs to happen quickly.

With the new systems, “the software finds the optimal adjustments for set points,” for example, changing voltages can allow for power to be redirected through less-congested lines, or can be increased to lessen power losses, Anton says.

The software also helps in the long-term planning for the grid. As many fossil-fuel power plants are scheduled to be decommissioned soon in Puerto Rico, as they are in many other places, planning for how to replace that power without having to resort to greenhouse gas-emitting sources is a key to achieving carbon-reduction goals. And by analyzing usage patterns, the software can guide the placement of new renewable power sources where they can most efficiently provide power where and when it’s needed.

As plants are retired or as components are affected by weather, “We wanted to ensure the dispatchability of power when the load changes,” Anton says, “but also when crucial components are lost, to ensure the robustness at each step of the retirement schedule.”

One thing they found was that “if you look at how much generating capacity exists, it’s more than the peak load, even after you retire a few fossil plants,” Ilic says. “But it’s hard to deliver.” Strategic planning of new distribution lines could make a big difference.

Jaddivada, director of innovation at SmartGridz, says that “we evaluated different possible architectures in Puerto Rico, and we showed the ability of this software to ensure uninterrupted electricity service. This is the most important challenge utilities have today. They have to go through a computationally tedious process to make sure the grid functions for any possible outage in the system. And that can be done in a much more efficient way through the software that the company  developed.”

The project was a collaborative effort between the MIT LIDS researchers and others at MIT Lincoln Laboratory, the Pacific Northwest National Laboratory, with overall help of SmartGridz software. 

Used in everything from soda cans and foil wrap to circuit boards and rocket boosters, aluminum is the second-most-produced metal in the world after steel. By the end of this decade, demand is projected to drive up aluminum production by 40 percent worldwide. This steep rise will magnify aluminum’s environmental impacts, including any pollutants that are released with its manufacturing waste.

MIT engineers have developed a new nanofiltration process to curb the hazardous waste generated from aluminum production. Nanofiltration could potentially be used to process the waste from an aluminum plant and retrieve any aluminum ions that would otherwise have escaped in the effluent stream. The captured aluminum could then be upcycled and added to the bulk of the produced aluminum, increasing yield while simultaneously reducing waste.

The researchers demonstrated the membrane’s performance in lab-scale experiments using a novel membrane to filter various solutions that were similar in content to the waste streams produced by aluminum plants. They found that the membrane selectively captured more than 99 percent of aluminum ions in these solutions.

If scaled up and implemented in existing production facilities, the membrane technology could reduce the amount of wasted aluminum and improve the environmental quality of the waste that plants generate.

“This membrane technology not only cuts down on hazardous waste but also enables a circular economy for aluminum by reducing the need for new mining,” says John Lienhard, the Abdul Latif Jameel Professor of Water in the Department of Mechanical Engineering, and director of the Abdul Latif Jameel Water and Food Systems Lab (J-WAFS) at MIT. “This offers a promising solution to address environmental concerns while meeting the growing demand for aluminum.”

Lienhard and his colleagues report their results in a study appearing today in the journal ACS Sustainable Chemistry and Engineering. The study’s co-authors include MIT mechanical engineering undergraduates Trent Lee and Vinn Nguyen, and Zi Hao Foo SM ’21, PhD ’24, who is a postdoc at the University of California at Berkeley.

A recycling niche

Lienhard’s group at MIT develops membrane and filtration technologies for desalinating seawater and remediating various sources of wastewater. In looking for new areas to apply their work, the team found an unexplored opportunity in aluminum and, in particular, the wastewater generated from the metal’s production.

As part of aluminum’s production, metal-rich ore, called bauxite, is first mined from open pits, then put through a series of chemical reactions to separate the aluminum from the rest of the mined rock. These reactions ultimately produce aluminum oxide, in a powdery form called alumina. Much of this alumina is then shipped to refineries, where the powder is poured into electrolysis vats containing a molten mineral called cryolite. When a strong electric current is applied, cryolite breaks alumina’s chemical bonds, separating aluminum and oxygen atoms. The pure aluminum then settles in liquid form to the bottom of the vat, where it can be collected and cast into various forms.

Cryolite electrolyte acts as a solvent, facilitating the separation of alumina during the molten salt electrolysis process. Over time, the cryolite accumulates impurities such as sodium, lithium, and potassium ions — gradually reducing its effectiveness in dissolving alumina. At a certain point, the concentration of these impurities reaches a critical level, at which the electrolyte must be replaced with fresh cryolite to main process efficiency. The spent cryolite, a viscous sludge containing residual aluminum ions and impurities, is then transported away for disposal.   

“We learned that for a traditional aluminum plant, something like 2,800 tons of aluminum are wasted per year,” says lead author Trent Lee, who carried out the new work as part of the MITEI Energy UROP program. “We were looking at ways that the industry can be more efficient, and we found cryolite waste hadn’t been well-researched in terms of recycling some of its waste products.”

A charged kick

In their new work, the researchers aimed to develop a membrane process to filter cryolite waste and recover aluminum ions that inevitably make it into the waste stream. Specifically, the team looked to capture aluminum while letting through all other ions, especially sodium, which builds up significantly in the cryolite over time.

The team reasoned that if they could selectively capture aluminum from cryolite waste, the aluminum could be poured back into the electrolysis vat without adding excessive sodium that would further slow the electrolysis process.

The researchers’ new design is an adaptation of membranes used in conventional water treatment plants. These membranes are typically made from a thin sheet of polymer material that is perforated by tiny, nanometer-scale pores, the size of which is tuned to let through specific ions and molecules.

The surface of conventional membranes carries a natural, negative charge. As a result, the membranes repel any ions that carry the same negative charge, while they attract positively charged ions to flow through.

In collaboration with the Japanese membrane company Nitto Denko, the MIT team sought to examine the efficacy of commercially available membranes that could filter through most positively charged ions in cryolite wastewater while repelling and capturing aluminum ions. However, aluminum ions also carry a positive charge, of +3, where sodium and the other cations carry a lesser positive charge of +1.

Motivated by the group’s recent work investigating membranes for recovering lithium from salt lakes and spent batteries, the team tested a novel Nitto Denko membrane with a thin, positively charged coating covering the membrane. The coating’s charge is just positive enough to strongly repel and retain aluminum while allowing less positively charged ions to flow through.

“The aluminum is the most positively charged of the ions, so most of it is kicked away from the membrane,” Foo explains.

The team tested the membrane’s performance by passing through solutions with various balances of ions, similar to what can be found in cryolite waste. They observed that the membrane consistently captured 99.5 percent of aluminum ions while allowing through sodium and the other cations. They also varied the pH of the solutions, and found the membrane maintained its performance even after sitting in highly acidic solution for several weeks.

“A lot of this cryolite waste stream comes at different levels of acidity,” Foo says. “And we found the membrane works really well, even within the harsh conditions that we would expect.”

The new experimental membrane is about the size of a playing card. To treat cryolite waste in an industrial-scale aluminum production plant, the researchers envision a scaled-up version of the membrane, similar to what is used in many desalination plants, where a long membrane is rolled up in a spiral configuration, through which water flows.

“This paper shows the viability of membranes for innovations in circular economies,” Lee says. “This membrane provides the dual benefit of upcycling aluminum while reducing hazardous waste.”

When a new species is introduced into an ecosystem, it may succeed in establishing itself, or it may fail to gain a foothold and die out. Physicists at MIT have now devised a formula that can predict which of those outcomes is most likely.

The researchers created their formula based on analysis of hundreds of different scenarios that they modeled using populations of soil bacteria grown in their laboratory. They now plan to test their formula in larger-scale ecosystems, including forests. This approach could also be helpful in predicting whether probiotics or fecal microbiota treatments (FMT) would successfully combat infections of the human GI tract.

“People eat a lot of probiotics, but many of them can never invade our gut microbiome at all, because if you introduce it, it does not necessarily mean that it can grow and colonize and benefit your health,” says Jiliang Hu SM ’19, PhD ’24, the lead author of the study.

MIT professor of physics Jeff Gore is the senior author of the paper, which appears today in the journal Nature Ecology and Evolution. Matthieu Barbier, a researcher at the Plant Health Institute Montpellier, and Guy Bunin, a professor of physics at Technion, are also authors of the paper.

Population fluctuations

Gore’s lab specializes in using microbes to analyze interspecies interactions in a controlled way, in hopes of learning more about how natural ecosystems behave. In previous work, the team has used bacterial populations to demonstrate how changing the environment in which the microbes live affects the stability of the communities they form.

In this study, the researchers wanted to study what determines whether an invasion by a new species will succeed or fail. In natural communities, ecologists have hypothesized that the more diverse an ecosystem is, the more it will resist an invasion, because most of the ecological niches will already be occupied and few resources are left for an invader.

However, in both natural and experimental systems, scientists have observed that this is not consistently true: While some highly diverse populations are resistant to invasion, other highly diverse populations are more likely to be invaded.

To explore why both of those outcomes can occur, the researchers set up more than 400 communities of soil bacteria, which were all native to the soil around MIT. The researchers established communities of 12 to 20 species of bacteria, and six days later, they added one randomly chosen species as the invader. On the 12th day of the experiment, they sequenced the genomes of all the bacteria to determine if the invader had established itself in the ecosystem.

In each community, the researchers also varied the nutrient levels in the culture medium on which the bacteria were grown. When nutrient levels were high, the microbes displayed strong interactions, characterized by heightened competition for food and other resources, or mutual inhibition through mechanisms such as pH-mediated cross-toxin effects. Some of these populations formed stable states in which the fraction of each microbe did not vary much over time, while others formed communities in which most of the species fluctuated in number.

The researchers found that these fluctuations were the most important factor in the outcome of the invasion. Communities that had more fluctuations tended to be more diverse, but they were also more likely to be invaded successfully.

“The fluctuation is not driven by changes in the environment, but it is internal fluctuation driven by the species interaction. And what we found is that the fluctuating communities are more readily invaded and also more diverse than the stable ones,” Hu says.

In some of the populations where the invader established itself, the other species remained, but in smaller numbers. In other populations, some of the resident species were outcompeted and disappeared completely. This displacement tended to happen more often in ecosystems when there were stronger competitive interactions between species.

In ecosystems that had more stable, less diverse populations, with stronger interactions between species, invasions were more likely to fail.

Regardless of whether the community was stable or fluctuating, the researchers found that the fraction of the original species that survived in the community before invasion predicts the probability of invasion success. This “survival fraction” could be estimated in natural communities by taking the ratio of the diversity within a local community (measured by the number of species in that area) to the regional diversity (number of species found in the entire region).

“It would be exciting to study whether the local and regional diversity could be used to predict susceptibility to invasion in natural communities,” Gore says.

Predicting success

The researchers also found that under certain circumstances, the order in which species arrived in the ecosystem played a role in whether an invasion was successful. When the interactions between species were strong, the chances of a species becoming successfully incorporated went down when that species was introduced after other species have already become established.

When the interactions are weak, this “priority effect” disappears and the same stable equilibrium is reached no matter what order the microbes arrived in.

“Under a strong interaction regime, we found the invader has some disadvantage because it arrived later. This is of interest in ecology because people have always found that in some cases the order in which species arrived matters a lot, while in the other cases it doesn't matter,” Hu says.

The researchers now plan to try to replicate their findings in ecosystems for which species diversity data is available, including the human gut microbiome. Their formula could allow them to predict the success of probiotic treatment, in which beneficial bacteria are consumed orally, or FMT, an experimental treatment for severe infections such as C. difficile, in which beneficial bacteria from a donor’s stool are transplanted into a patient’s colon.

“Invasions can be harmful or can be good depending on the context,” Hu says. “In some cases, like probiotics, or FMT to treat C. difficile infection, we want the healthy species to invade successfully. Also for soil protection, people introduce probiotics or beneficial species to the soil. In that case people also want the invaders to succeed.”

The research was funded by the Schmidt Polymath Award and the Sloan Foundation.

One of the hardest-working organisms in the ocean is the tiny, emerald-tinged Prochlorococcus marinus. These single-celled “picoplankton,” which are smaller than a human red blood cell, can be found in staggering numbers throughout the ocean’s surface waters, making Prochlorococcus the most abundant photosynthesizing organism on the planet. (Collectively, Prochlorococcus fix as much carbon as all the crops on land.) Scientists continue to find new ways that the little green microbe is involved in the ocean’s cycling and storage of carbon.

Now, MIT scientists have discovered a new ocean-regulating ability in the small but mighty microbes: cross-feeding of DNA building blocks. In a study appearing today in Science Advances, the team reports that Prochlorococcus shed these extra compounds into their surroundings, where they are then “cross-fed,” or taken up by other ocean organisms, either as nutrients, energy, or for regulating metabolism. Prochlorococcus’ rejects, then, are other microbes’ resources.

What’s more, this cross-feeding occurs on a regular cycle: Prochlorococcus tend to shed their molecular baggage at night, when enterprising microbes quickly consume the cast-offs. For a microbe called SAR11, the most abundant bacteria in the ocean, the researchers found that the nighttime snack acts as a relaxant of sorts, forcing the bacteria to slow down their metabolism and effectively recharge for the next day.

Through this cross-feeding interaction, Prochlorococcus could be helping many microbial communities to grow sustainably, simply by giving away what it doesn’t need. And they’re doing so in a way that could set the daily rhythms of microbes around the world.

“The relationship between the two most abundant groups of microbes in ocean ecosystems has intrigued oceanographers for years,” says co-author and MIT Institute Professor Sallie “Penny” Chisholm, who played a role in the discovery of Prochlorococcus in 1986. “Now we have a glimpse of the finely tuned choreography that contributes to their growth and stability across vast regions of the oceans.”

Given that Prochlorococcus and SAR11 suffuse the surface oceans, the team suspects that the exchange of molecules from one to the other could amount to one of the major cross-feeding relationships in the ocean, making it an important regulator of the ocean carbon cycle.

“By looking at the details and diversity of cross-feeding processes, we can start to unearth important forces that are shaping the carbon cycle,” says the study’s lead author, Rogier Braakman, a research scientist in MIT’s Department of Earth, Atmospheric and Planetary Sciences (EAPS).

Other MIT co-authors include Brandon Satinsky, Tyler O’Keefe, Shane Hogle, Jamie Becker, Robert Li, Keven Dooley, and Aldo Arellano, along with Krista Longnecker, Melissa Soule, and Elizabeth Kujawinski of Woods Hole Oceanographic Institution (WHOI).

Spotting castaways

Cross-feeding occurs throughout the microbial world, though the process has mainly been studied in close-knit communities. In the human gut, for instance, microbes are in close proximity and can easily exchange and benefit from shared resources.

By comparison, Prochlorococcus are free-floating microbes that are regularly tossed and mixed through the ocean’s surface layers. While scientists assume that the plankton are involved in some amount of cross-feeding, exactly how this occurs, and who would benefit, have historically been challenging to probe; any stuff that Prochlorococcus cast away would have vanishingly low concentrations,and be exceedingly difficult to measure.

But in work published in 2023, Braakman teamed up with scientists at WHOI, who pioneered ways to measure small organic compounds in seawater. In the lab, they grew various strains of Prochlorococcus under different conditions and characterized what the microbes released. They found that among the major “exudants,” or released molecules, were purines and pyridines, which are molecular building blocks of DNA. The molecules also happen to be nitrogen-rich — a fact that puzzled the team. Prochlorococcus are mainly found in ocean regions that are low in nitrogen, so it was assumed they’d want to retain any and all nitrogen-containing compounds they can. Why, then, were they instead throwing such compounds away?

Global symphony

In their new study, the researchers took a deep dive into the details of Prochlorococcus’ cross-feeding and how it influences various types of ocean microbes.

They set out to study how Prochlorococcus use purine and pyridine in the first place, before expelling the compounds into their surroundings. They compared published genomes of the microbes, looking for genes that encode purine and pyridine metabolism. Tracing the genes forward through the genomes, the team found that once the compounds are produced, they are used to make DNA and replicate the microbes’ genome. Any leftover purine and pyridine is recycled and used again, though a fraction of the stuff is ultimately released into the environment. Prochlorococcus appear to make the most of the compounds, then cast off what they can’t.

The team also looked to gene expression data and found that genes involved in recycling purine and pyrimidine peak several hours after the recognized peak in genome replication that occurs at dusk. The question then was: What could be benefiting from this nightly shedding?

For this, the team looked at the genomes of more than 300 heterotrophic microbes — organisms that consume organic carbon rather than making it themselves through photosynthesis. They suspected that such carbon-feeders could be likely consumers of Prochlorococcus’ organic rejects. They found most of the heterotrophs contained genes that take up either purine or pyridine, or in some cases, both, suggesting microbes have evolved along different paths in terms of how they cross-feed.

The group zeroed in on one purine-preferring microbe, SAR11, as it is the most abundant heterotrophic microbe in the ocean. When they then compared the genes across different strains of SAR11, they found that various types use purines for different purposes, from simply taking them up and using them intact to breaking them down for their energy, carbon, or nitrogen. What could explain the diversity in how the microbes were using Prochlorococcus’ cast-offs?

It turns out the local environment plays a big role. Braakman and his collaborators performed a metagenome analysis in which they compared the collectively sequenced genomes of all microbes in over 600 seawater samples from around the world, focusing on SAR11 bacteria. Metagenome sequences were collected alongside measurements of various environmental conditions and geographic locations in which they are found. This analysis showed that the bacteria gobble up purine for its nitrogen when the nitrogen in seawater is low, and for its carbon or energy when nitrogen is in surplus — revealing the selective pressures shaping these communities in different ocean regimes.

“The work here suggests that microbes in the ocean have developed relationships that advance their growth potential in ways we don’t expect,” says co-author Kujawinski.

Finally, the team carried out a simple experiment in the lab, to see if they could directly observe a mechanism by which purine acts on SAR11. They grew the bacteria in cultures, exposed them to various concentrations of purine, and unexpectedly found it causes them to slow down their normal metabolic activities and even growth. However, when the researchers put these same cells under environmentally stressful conditions, they continued growing strong and healthy cells, as if the metabolic pausing by purines helped prime them for growth, thereby avoiding the effects of the stress.

“When you think about the ocean, where you see this daily pulse of purines being released by Prochlorococcus, this provides a daily inhibition signal that could be causing a pause in SAR11 metabolism, so that the next day when the sun comes out, they are primed and ready,” Braakman says. “So we think Prochlorococcus is acting as a conductor in the daily symphony of ocean metabolism, and cross-feeding is creating a global synchronization among all these microbial cells.”

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

By adapting artificial intelligence models known as large language models, researchers have made great progress in their ability to predict a protein’s structure from its sequence. However, this approach hasn’t been as successful for antibodies, in part because of the hypervariability seen in this type of protein.

To overcome that limitation, MIT researchers have developed a computational technique that allows large language models to predict antibody structures more accurately. Their work could enable researchers to sift through millions of possible antibodies to identify those that could be used to treat SARS-CoV-2 and other infectious diseases.

“Our method allows us to scale, whereas others do not, to the point where we can actually find a few needles in the haystack,” says Bonnie Berger, the Simons Professor of Mathematics, the head of the Computation and Biology group in MIT’s Computer Science and Artificial Intelligence Laboratory (CSAIL), and one of the senior authors of the new study. “If we could help to stop drug companies from going into clinical trials with the wrong thing, it would really save a lot of money.”

The technique, which focuses on modeling the hypervariable regions of antibodies, also holds potential for analyzing entire antibody repertoires from individual people. This could be useful for studying the immune response of people who are super responders to diseases such as HIV, to help figure out why their antibodies fend off the virus so effectively.

Bryan Bryson, an associate professor of biological engineering at MIT and a member of the Ragon Institute of MGH, MIT, and Harvard, is also a senior author of the paper, which appears this week in the Proceedings of the National Academy of Sciences. Rohit Singh, a former CSAIL research scientist who is now an assistant professor of biostatistics and bioinformatics and cell biology at Duke University, and Chiho Im ’22 are the lead authors of the paper. Researchers from Sanofi and ETH Zurich also contributed to the research.

Modeling hypervariability

Proteins consist of long chains of amino acids, which can fold into an enormous number of possible structures. In recent years, predicting these structures has become much easier to do, using artificial intelligence programs such as AlphaFold. Many of these programs, such as ESMFold and OmegaFold, are based on large language models, which were originally developed to analyze vast amounts of text, allowing them to learn to predict the next word in a sequence. This same approach can work for protein sequences — by learning which protein structures are most likely to be formed from different patterns of amino acids.

However, this technique doesn’t always work on antibodies, especially on a segment of the antibody known as the hypervariable region. Antibodies usually have a Y-shaped structure, and these hypervariable regions are located in the tips of the Y, where they detect and bind to foreign proteins, also known as antigens. The bottom part of the Y provides structural support and helps antibodies to interact with immune cells.

Hypervariable regions vary in length but usually contain fewer than 40 amino acids. It has been estimated that the human immune system can produce up to 1 quintillion different antibodies by changing the sequence of these amino acids, helping to ensure that the body can respond to a huge variety of potential antigens. Those sequences aren’t evolutionarily constrained the same way that other protein sequences are, so it’s difficult for large language models to learn to predict their structures accurately.

“Part of the reason why language models can predict protein structure well is that evolution constrains these sequences in ways in which the model can decipher what those constraints would have meant,” Singh says. “It’s similar to learning the rules of grammar by looking at the context of words in a sentence, allowing you to figure out what it means.”

To model those hypervariable regions, the researchers created two modules that build on existing protein language models. One of these modules was trained on hypervariable sequences from about 3,000 antibody structures found in the Protein Data Bank (PDB), allowing it to learn which sequences tend to generate similar structures. The other module was trained on data that correlates about 3,700 antibody sequences to how strongly they bind three different antigens.

The resulting computational model, known as AbMap, can predict antibody structures and binding strength based on their amino acid sequences. To demonstrate the usefulness of this model, the researchers used it to predict antibody structures that would strongly neutralize the spike protein of the SARS-CoV-2 virus.

The researchers started with a set of antibodies that had been predicted to bind to this target, then generated millions of variants by changing the hypervariable regions. Their model was able to identify antibody structures that would be the most successful, much more accurately than traditional protein-structure models based on large language models.

Then, the researchers took the additional step of clustering the antibodies into groups that had similar structures. They chose antibodies from each of these clusters to test experimentally, working with researchers at Sanofi. Those experiments found that 82 percent of these antibodies had better binding strength than the original antibodies that went into the model.

Identifying a variety of good candidates early in the development process could help drug companies avoid spending a lot of money on testing candidates that end up failing later on, the researchers say.

“They don’t want to put all their eggs in one basket,” Singh says. “They don’t want to say, I’m going to take this one antibody and take it through preclinical trials, and then it turns out to be toxic. They would rather have a set of good possibilities and move all of them through, so that they have some choices if one goes wrong.”

Comparing antibodies

Using this technique, researchers could also try to answer some longstanding questions about why different people respond to infection differently. For example, why do some people develop much more severe forms of Covid, and why do some people who are exposed to HIV never become infected?

Scientists have been trying to answer those questions by performing single-cell RNA sequencing of immune cells from individuals and comparing them — a process known as antibody repertoire analysis. Previous work has shown that antibody repertoires from two different people may overlap as little as 10 percent. However, sequencing doesn’t offer as comprehensive a picture of antibody performance as structural information, because two antibodies that have different sequences may have similar structures and functions.

The new model can help to solve that problem by quickly generating structures for all of the antibodies found in an individual. In this study, the researchers showed that when structure is taken into account, there is much more overlap between individuals than the 10 percent seen in sequence comparisons. They now plan to further investigate how these structures may contribute to the body’s overall immune response against a particular pathogen.

“This is where a language model fits in very beautifully because it has the scalability of sequence-based analysis, but it approaches the accuracy of structure-based analysis,” Singh says.

The research was funded by Sanofi and the Abdul Latif Jameel Clinic for Machine Learning in Health. 

Fast radio bursts are brief and brilliant explosions of radio waves emitted by extremely compact objects such as neutron stars and possibly black holes. These fleeting fireworks last for just a thousandth of a second and can carry an enormous amount of energy — enough to briefly outshine entire galaxies.

Since the first fast radio burst (FRB) was discovered in 2007, astronomers have detected thousands of FRBs, whose locations range from within our own galaxy to as far as 8 billion light-years away. Exactly how these cosmic radio flares are launched is a highly contested unknown.

Now, astronomers at MIT have pinned down the origins of at least one fast radio burst using a novel technique that could do the same for other FRBs. In their new study, appearing today in the journal Nature, the team focused on FRB 20221022A — a previously discovered fast radio burst that was detected from a galaxy about 200 million light-years away.

The team zeroed in further to determine the precise location of the radio signal by analyzing its “scintillation,” similar to how stars twinkle in the night sky. The scientists studied changes in the FRB’s brightness and determined that the burst must have originated from the immediate vicinity of its source, rather than much further out, as some models have predicted.

The team estimates that FRB 20221022A exploded from a region that is extremely close to a rotating neutron star, 10,000 kilometers away at most. That’s less than the distance between New York and Singapore. At such close range, the burst likely emerged from the neutron star’s magnetosphere — a highly magnetic region immediately surrounding the ultracompact star.

The team’s findings provide the first conclusive evidence that a fast radio burst can originate from the magnetosphere, the highly magnetic environment immediately surrounding an extremely compact object.

“In these environments of neutron stars, the magnetic fields are really at the limits of what the universe can produce,” says lead author Kenzie Nimmo, a postdoc in MIT’s Kavli Institute for Astrophysics and Space Research. “There’s been a lot of debate about whether this bright radio emission could even escape from that extreme plasma.”

“Around these highly magnetic neutron stars, also known as magnetars, atoms can’t exist — they would just get torn apart by the magnetic fields,” says Kiyoshi Masui, associate professor of physics at MIT. “The exciting thing here is, we find that the energy stored in those magnetic fields, close to the source, is twisting and reconfiguring such that it can be released as radio waves that we can see halfway across the universe.”

The study’s MIT co-authors include Adam Lanman, Shion Andrew, Daniele Michilli, and Kaitlyn Shin, along with collaborators from multiple institutions.

Burst size

Detections of fast radio bursts have ramped up in recent years, due to the Canadian Hydrogen Intensity Mapping Experiment (CHIME). The radio telescope array comprises four large, stationary receivers, each shaped like a half-pipe, that are tuned to detect radio emissions within a range that is highly sensitive to fast radio bursts.

Since 2020, CHIME has detected thousands of FRBs from all over the universe. While scientists generally agree that the bursts arise from extremely compact objects, the exact physics driving the FRBs is unclear. Some models predict that fast radio bursts should come from the turbulent magnetosphere immediately surrounding a compact object, while others predict that the bursts should originate much further out, as part of a shockwave that propagates away from the central object.

To distinguish between the two scenarios, and determine where fast radio bursts arise, the team considered scintillation — the effect that occurs when light from a small bright source such as a star, filters through some medium, such as a galaxy’s gas. As the starlight filters through the gas, it bends in ways that make it appear, to a distant observer, as if the star is twinkling. The smaller or the farther away an object is, the more it twinkles. The light from larger or closer objects, such as planets in our own solar system, experience less bending, and therefore do not appear to twinkle.

The team reasoned that if they could estimate the degree to which an FRB scintillates, they might determine the relative size of the region from where the FRB originated. The smaller the region, the closer in the burst would be to its source, and the more likely it is to have come from a magnetically turbulent environment. The larger the region, the farther the burst would be, giving support to the idea that FRBs stem from far-out shockwaves.

Twinkle pattern

To test their idea, the researchers looked to FRB 20221022A, a fast radio burst that was detected by CHIME in 2022. The signal lasts about two milliseconds, and is a relatively run-of-the-mill FRB, in terms of its brightness. However, the team’s collaborators at McGill University found that FRB 20221022A exhibited one standout property: The light from the burst was highly polarized, with the angle of polarization tracing a smooth S-shaped curve.  This pattern is interpreted as evidence that the FRB emission site is rotating — a characteristic previously observed in pulsars, which are highly magnetized, rotating neutron stars.

To see a similar polarization in fast radio bursts was a first, suggesting that the signal may have arisen from the close-in vicinity of a neutron star. The McGill team’s results are reported in a companion paper today in Nature.

The MIT team realized that if FRB 20221022A originated from close to a neutron star, they should be able to prove this, using scintillation.

In their new study, Nimmo and her colleagues analyzed data from CHIME and observed steep variations in brightness that signaled scintillation — in other words, the FRB was twinkling. They confirmed that there is gas somewhere between the telescope and FRB that is bending and filtering the radio waves. The team then determined where this gas could be located, confirming that gas within the FRB’s host galaxy was responsible for some of the scintillation observed. This gas acted as a natural lens, allowing the researchers to zoom in on the FRB site and determine that the burst originated from an extremely small region, estimated to be about 10,000 kilometers wide.

“This means that the FRB is probably within hundreds of thousands of kilometers from the source,” Nimmo says. “That’s very close. For comparison, we would expect the signal would be more than tens of millions of kilometers away if it originated from a shockwave, and we would see no scintillation at all.”

“Zooming in to a 10,000-kilometer region, from a distance of 200 million light years, is like being able to measure the width of a DNA helix, which is about 2 nanometers wide, on the surface of the moon,” Masui says. “There’s an amazing range of scales involved.”

The team’s results, combined with the findings from the McGill team, rule out the possibility that FRB 20221022A emerged from the outskirts of a compact object. Instead, the studies prove for the first time that fast radio bursts can originate from very close to a neutron star, in highly chaotic magnetic environments.

“These bursts are always happening, and CHIME detects several a day,” Masui says. “There may be a lot of diversity in how and where they occur, and this scintillation technique will be really useful in helping to disentangle the various physics that drive these bursts.”

“The pattern traced by the polarization angle was so strikingly similar to that seen from pulsars in our own Milky Way Galaxy that there was some initial concern that the source wasn't actually an FRB but a misclassified pulsar,” says Ryan Mckinven, a co-author of the study from McGill University. “Fortunately, these concerns were put to rest with the help of data collected from an optical telescope that confirmed the FRB originated in a galaxy millions of light-years away.”

“Polarimetry is one of the few tools we have to probe these distant sources,” Mckinven explains. “This result will likely inspire follow-up studies of similar behavior in other FRBs and prompt theoretical efforts to reconcile the differences in their polarized signals.”

This research was supported by various institutions including the Canada Foundation for Innovation, the Dunlap Institute for Astronomy and Astrophysics at the University of Toronto, the Canadian Institute for Advanced Research, the Trottier Space Institute at McGill University, and the University of British Columbia.

MIT’s research community had another year full of scientific and technological advances in 2024. To celebrate the achievements of the past twelve months, MIT News highlights some of our most popular stories from this year. We’ve also rounded up the year’s top MIT community-related stories.

• 3-D printing with liquid metal: Researchers developed an additive manufacturing technique that can print rapidly with liquid metal, producing large-scale parts like table legs and chair frames in a matter of minutes. Their technique involves depositing molten aluminum along a predefined path into a bed of tiny glass beads. The aluminum quickly hardens into a 3D structure.
   
• Tamper-proof ID tags: Engineers developed a tag that can reveal with near-perfect accuracy whether an item is real or fake. The key is in the glue that sticks the tag to the item. The team uses terahertz waves to authenticate items by recognizing a unique pattern of microscopic metal particles mixed into the glue.
   
• Chatting with the future you: 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. The project is aimed at reducing anxiety and guiding young people to make better choices.
   
• Converting CO2 into useful products: Engineers at MIT designed a new electrode that boosts the efficiency of electrochemical reactions to turn carbon dioxide into ethylene and other products.
   
• Generative AI for databases: Researchers built GenSQL, a new generative AI tool that 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. The tool could help users make predictions, detect anomalies, guess missing values, fix errors, and more.
   
• Reversing autoimmune-induced hair loss: A new microneedle patch delivers immune-regulating molecules to the scalp. The treatment teaches T cells not to attack hair follicles, promoting hair regrowth and offering a promising solution for individuals affected by alopecia areata and other autoimmune skin diseases.
   
• Inside the LLM black box: Researchers demonstrated a technique that can be used to probe a large language model to see what it knows about new subjects. The technique showed the models use a surprisingly simple mechanism to retrieve some stored knowledge.
   
• Sound-suppressing silk: An interdisciplinary collaboration of researchers from MIT and elsewhere developed a silk fabric, barely thicker than a human hair, that can suppress unwanted noise and reduce noise transmission in a large room.
   
• Working out for your nervous system: Researchers found that when muscles work out, they help neurons to grow as well. The findings suggest that biochemical and physical effects of exercise could help heal nerves.
   
• Finding AI’s world model lacking: Researchers found that despite its impressive output, generative AI models don’t have a coherent understanding of the world. Large language models don't form true models of the world and its rules, and can thus fail unexpectedly on similar tasks.

Within the human digestive tract are trillions of bacteria from thousands of different species. These bacteria form communities that help digest food, fend off harmful microbes, and play many other roles in maintaining human health.

These bacteria can be vulnerable to infection from viruses called bacteriophages. One of bacterial cells’ most well-known defenses against these viruses is the CRISPR system, which evolved in bacteria to help them recognize and chop up viral DNA.

A study from MIT biological engineers has yielded new insight into how bacteria in the gut microbiome adapt their CRISPR defenses as they encounter new threats. The researchers found that while bacteria grown in the lab can incorporate new viral recognition sequences as quickly as once a day, bacteria living in human gut add new sequences at a much slower rate — on average, one every three years.

The findings suggest that the environment within the digestive tract offers many fewer opportunities for bacteria and bacteriophages to interact than in the lab, so bacteria don’t need to update their CRISPR defenses very often. It also raises the question of whether bacteria have more important defense systems than CRISPR.

“This finding is significant because we use microbiome-based therapies like fecal microbiota transplant to help treat some diseases, but efficacy is inconsistent because new microbes do not always survive in patients. Learning about microbial defenses against viruses helps us to understand what makes a strong, healthy microbial community,” says An-Ni Zhang, a former MIT postdoc who is now an assistant professor at Nanyang Technological University.

Zhang is the lead author of the study, which appears today in the journal Cell Genomics. Eric Alm, director of MIT’s Center for Microbiome Informatics and Therapeutics, a professor of biological engineering and of civil and environmental engineering at MIT, and a member of the Broad Institute of MIT and Harvard, is the paper’s senior author.

Infrequent exposure

In bacteria, CRISPR serves as a memory immune response. When bacteria encounter viral DNA, they can incorporate part of the sequence into their own DNA. Then, if the virus is encountered again, that sequence produces a guide RNA that directs an enzyme called Cas9 to snip the viral DNA, preventing infection.

These virus-specific sequences are called spacers, and a single bacterial cell may carry more than 200 spacers. These sequences can be passed onto offspring, and they can also be shared with other bacterial cells through a process called horizontal gene transfer.

Previous studies have found that spacer acquisition occurs very rapidly in the lab, but the process appears to be slower in natural environments. In the new study, the MIT team wanted to explore how often this process happens in bacteria in the human gut.

“We were interested in how fast this CRISPR system changes its spacers, specifically in the gut microbiome, to better understand the bacteria-virus interactions inside our body,” Zhang says. “We wanted to identify the key parameters that impact the timescale of this immunity update.”

To do that, the researchers looked at how CRISPR sequences changed over time in two different datasets obtained by sequencing microbes from the human digestive tract. One of these datasets contained 6,275 genomic sequences representing 52 bacterial species, and the other contained 388 longitudinal “metagenomes,” that is, sequences from many microbes found in a sample, taken from four healthy people.

“By analyzing those two datasets, we found out that spacer acquisition is really slow in human gut microbiome: On average, it would take 2.7 to 2.9 years for a bacterial species to acquire a single spacer in our gut, which is super surprising because our gut is challenged with viruses almost every day from the microbiome itself and in our food,” Zhang says.

The researchers then built a computational model to help them figure out why the acquisition rate was so slow. This analysis showed that spacers are acquired more rapidly when bacteria live in high-density populations. However, the human digestive tract is diluted several times a day, whenever a meal is consumed. This flushes out some bacteria and viruses and keeps the overall density low, making it less likely that the microbes will encounter a virus that can infect them.

Another factor may be the spatial distribution of microbes, which the researchers believe prevents some bacteria from encountering viruses very frequently.

“Sometimes one population of bacteria may never or rarely encounter a phage because the bacteria are closer to the epithelium in the mucus layer and farther away from a potential exposure to viruses,” Zhang says.

Bacterial interactions

Among the populations of bacteria that they studied, the researchers identified one species — Bifidobacteria longum — that had gained spacers much more recently than others. The researchers found that in samples from unrelated people, living on different continents, B. longum had recently acquired up to six different spacers targeting two different Bifidobacteria bacteriophages.

This acquisition was driven by horizontal gene transfer — a process that allows bacteria to gain new genetic material from their neighbors. The findings suggest that there may be evolutionary pressure on B. longum from those two viruses.

“It has been highly overlooked how much horizontal gene transfer contributes to this dynamic. Within communities of bacteria, the bacteria-bacteria interactions can be a main contributor to the development of viral resistance,” Zhang says.

Analyzing microbes’ immune defenses may offer a way for scientists to develop targeted treatments that will be most effective in a particular patient, the researchers say. For example, they could design therapeutic microbes that are able to fend off the types of bacteriophages that are most prevalent in that person’s microbiome, which would increase the chances that the treatment would succeed.

“One thing we can do is to study the viral composition in the patients, and then we can identify which microbiome species or strains are more capable of resisting those local viruses in a person,” Zhang says.

The research was funded, in part, by the Broad Institute and the Thomas and Stacey Siebel Foundation.

Imagine that the head of a company office is misbehaving, and a disillusioned employee reports the problem to their manager. Instead of the complaint getting traction, however, the manager sidesteps the issue and implies that raising it further could land the unhappy employee in trouble — but doesn’t deny that the problem exists.

This hypothetical scenario involves an open secret: a piece of information that is widely known but never acknowledged as such. Open secrets often create practical quandaries for people, as well as backlash against those who try to address the things that the secrets protect.

In a newly published paper, MIT philosopher Sam Berstler contends that open secrets are pervasive and problematic enough to be worthy of systematic study — and provides a detailed analysis of the distinctive social dynamics accompanying them. In many cases, she proposes, ignoring some things is fine — but open secrets present a special problem.

After all, people might maintain friendships better by not disclosing their salaries to each other, and relatives might get along better if they avoid talking politics at the holidays. But these are just run-of-the-mill individual decisions.

By contrast, open secrets are especially damaging, Berstler believes, because of their “iterative” structure. We do not talk about open secrets; we do not talk about the fact that we do not talk about them; and so on, until the possibility of addressing the problems at hand disappears.

“Sometimes not acknowledging things can be very productive,” Berstler says. “It’s good we don’t talk about everything in the workplace. What’s different about open secrecy is not the content of what we’re not acknowledging, but the pernicious iterative structure of our practice of not acknowledging it.  And because of that structure, open secrecy tends to be hard to change.”

Or, as she writes in the paper, “Open secrecy norms are often moral disasters.”

Beyond that, Berstler says, the example of open secrets should enable us to examine the nature of conversation itself in more multidimensional terms; we need to think about the things left unsaid in conversation, too.

Berstler’s paper, “The Structure of Open Secrets,” appears in advance online form in Philosophical Review. Berstler, an assistant professor and the Laurance S. Rockefeller Career Development Chair in MIT’s Department of Linguistics and Philosophy, is the sole author.

Eroding our knowledge

The concept of open secrets is hardly new, but it has not been subject to extensive philosophical rigor. The German sociologist Georg Simmel wrote about them in the early 20th century, but mostly in the context of secret societies keeping quirky rituals to themselves. Other prominent thinkers have addressed open secrets in psychological terms. To Berstler, the social dynamics of open secrets merit a more thorough reckoning.

“It’s not a psychological problem that people are having,” she says. “It’s a particular practice that they’re all conforming to. But it’s hard to see this because it’s the kind of practice that members, just in virtue of conforming to the practice, can’t talk about.”

In Berstler’s view, the iterative nature of open secrets distinguishes them. The employee expecting a candid reply from their manager may feel bewildered about the lack of a transparent response, and that nonacknowledgement means there is not much recourse to be had, either. Eventually, keeping open secrets means the original issue itself can be lost from view.

“Open secrets norms are set up to try to erode our knowledge,” Berstler says.

In practical terms, people may avoid addressing open secrets head-on because they face a familiar quandary: Being a whistleblower can cost people their jobs and more. But Berstler suggests in the paper that keeping open secrets helps people define their in-group status, too.

“It’s also the basis for group identity,” she says.

Berstler avoids taking the position that greater transparency is automatically a beneficial thing. The paper identifies at least one kind of special case where keeping open secrets might be good. Suppose, for instance, a co-worker has an eccentric but harmless habit their colleagues find out about: It might be gracious to spare them simple embarrassment.

That aside, as Berstler writes, open secrets “can serve as shields for powerful people guilty of serious, even criminal wrongdoing. The norms can compound the harm that befalls their victims … [who] find they don’t just have to contend with the perpetrator’s financial resources, political might, and interpersonal capital. They must go up against an entire social arrangement.” As a result, the chances of fixing social or organizational dysfunction diminish.

Two layers of conversation

Berstler is not only trying to chart the dynamics and problems of open secrets. She is also trying to usefully complicate our ideas about the nature of conversations and communication.

Broadly, some philosophers have theorized about conversations and communication by focusing largely on the information being shared among people. To Berstler, this is not quite sufficient; the example of open secrets alerts us that communication is not just an act of making things more and more transparent.

“What I’m arguing in the paper is that this is too simplistic a way to think about it, because actual conversations in the real world have a theatrical or dramatic structure,” Berstler says. “There are things that cannot be made explicit without ruining the performance.”

At an office holiday party, for instance, the company CEO might maintain an illusion of being on equal footing with the rest of the employees if the conversation is restricted to movies and television shows. If the subject turns to year-end bonuses, that illusion vanishes. Or two friends at a party, trapped in an unwanted conversation with a third person, might maneuver themselves away with knowing comments, but without explicitly saying they are trying to end the chat.

Here Berstler draws upon the work of sociologist Erving Goffman — who closely studied the performative aspects of everyday behavior — to outline how a more multi-dimensional conception of social interaction applies to open secrets. Berstler suggests open secrets involve what she calls “activity layering,” which in this case suggests that people in a conversation involving open secrets have multiple common grounds for understanding, but some remain unspoken.

Further expanding on Goffman’s work, Berstler also details how people may be “mutually collaborating on a pretense,” as she writes, to keep an open secret going.

“Goffman has not really systematically been brought into the philosophy of language, so I am showing how his ideas illuminate and complicate philosophical views,” Berstler says.

Combined, a close analysis of open secrets and a re-evaluation of the performative components of conversation can help us become more cognizant about communication. What is being said matters; what is left unsaid matters alongside it.

“There are structural features of open secrets that are worrisome,” Berstler says. “And because of that we have to more aware [of how they work].”

Try taking a picture of each of North America's roughly 11,000 tree species, and you’ll have a mere fraction of the millions of photos within nature image datasets. These massive collections of snapshots — ranging from butterflies to humpback whales — are a great research tool for ecologists because they provide evidence of organisms’ unique behaviors, rare conditions, migration patterns, and responses to pollution and other forms of climate change.

While comprehensive, nature image datasets aren’t yet as useful as they could be. It’s time-consuming to search these databases and retrieve the images most relevant to your hypothesis. You’d be better off with an automated research assistant — or perhaps artificial intelligence systems called multimodal vision language models (VLMs). They’re trained on both text and images, making it easier for them to pinpoint finer details, like the specific trees in the background of a photo.

But just how well can VLMs assist nature researchers with image retrieval? A team from MIT’s Computer Science and Artificial Intelligence Laboratory (CSAIL), University College London, iNaturalist, and elsewhere designed a performance test to find out. Each VLM’s task: locate and reorganize the most relevant results within the team’s “INQUIRE” dataset, composed of 5 million wildlife pictures and 250 search prompts from ecologists and other biodiversity experts. 

Looking for that special frog

In these evaluations, the researchers found that larger, more advanced VLMs, which are trained on far more data, can sometimes get researchers the results they want to see. The models performed reasonably well on straightforward queries about visual content, like identifying debris on a reef, but struggled significantly with queries requiring expert knowledge, like identifying specific biological conditions or behaviors. For example, VLMs somewhat easily uncovered examples of jellyfish on the beach, but struggled with more technical prompts like “axanthism in a green frog,” a condition that limits their ability to make their skin yellow.

Their findings indicate that the models need much more domain-specific training data to process difficult queries. MIT PhD student Edward Vendrow, a CSAIL affiliate who co-led work on the dataset in a new paper, believes that by familiarizing with more informative data, the VLMs could one day be great research assistants. “We want to build retrieval systems that find the exact results scientists seek when monitoring biodiversity and analyzing climate change,” says Vendrow. “Multimodal models don’t quite understand more complex scientific language yet, but we believe that INQUIRE will be an important benchmark for tracking how they improve in comprehending scientific terminology and ultimately helping researchers automatically find the exact images they need.”

The team’s experiments illustrated that larger models tended to be more effective for both simpler and more intricate searches due to their expansive training data. They first used the INQUIRE dataset to test if VLMs could narrow a pool of 5 million images to the top 100 most-relevant results (also known as “ranking”). For straightforward search queries like “a reef with manmade structures and debris,” relatively large models like “SigLIP” found matching images, while smaller-sized CLIP models struggled. According to Vendrow, larger VLMs are “only starting to be useful” at ranking tougher queries.

Vendrow and his colleagues also evaluated how well multimodal models could re-rank those 100 results, reorganizing which images were most pertinent to a search. In these tests, even huge LLMs trained on more curated data, like GPT-4o, struggled: Its precision score was only 59.6 percent, the highest score achieved by any model.

The researchers presented these results at the Conference on Neural Information Processing Systems (NeurIPS) earlier this month.

Inquiring for INQUIRE

The INQUIRE dataset includes search queries based on discussions with ecologists, biologists, oceanographers, and other experts about the types of images they’d look for, including animals’ unique physical conditions and behaviors. A team of annotators then spent 180 hours searching the iNaturalist dataset with these prompts, carefully combing through roughly 200,000 results to label 33,000 matches that fit the prompts.

For instance, the annotators used queries like “a hermit crab using plastic waste as its shell” and “a California condor tagged with a green ‘26’” to identify the subsets of the larger image dataset that depict these specific, rare events.

Then, the researchers used the same search queries to see how well VLMs could retrieve iNaturalist images. The annotators’ labels revealed when the models struggled to understand scientists’ keywords, as their results included images previously tagged as irrelevant to the search. For example, VLMs’ results for “redwood trees with fire scars” sometimes included images of trees without any markings.

“This is a careful curation of data, with a focus on capturing real examples of scientific inquiries across research areas in ecology and environmental science,” says Sara Beery, the Homer A. Burnell Career Development Assistant Professor at MIT, CSAIL principal investigator, and co-senior author of the work. “It’s proved vital to expanding our understanding of the current capabilities of VLMs in these potentially impactful scientific settings. It has also outlined gaps in current research that we can now work to address, particularly for complex compositional queries, technical terminology, and the fine-grained, subtle differences that delineate categories of interest for our collaborators.”

“Our findings imply that some vision models are already precise enough to aid wildlife scientists with retrieving some images, but many tasks are still too difficult for even the largest, best-performing models,” says Vendrow. “Although INQUIRE is focused on ecology and biodiversity monitoring, the wide variety of its queries means that VLMs that perform well on INQUIRE are likely to excel at analyzing large image collections in other observation-intensive fields.”

Inquiring minds want to see

Taking their project further, the researchers are working with iNaturalist to develop a query system to better help scientists and other curious minds find the images they actually want to see. Their working demo allows users to filter searches by species, enabling quicker discovery of relevant results like, say, the diverse eye colors of cats. Vendrow and co-lead author Omiros Pantazis, who recently received his PhD from University College London, also aim to improve the re-ranking system by augmenting current models to provide better results.

University of Pittsburgh Associate Professor Justin Kitzes highlights INQUIRE’s ability to uncover secondary data. “Biodiversity datasets are rapidly becoming too large for any individual scientist to review,” says Kitzes, who wasn’t involved in the research. “This paper draws attention to a difficult and unsolved problem, which is how to effectively search through such data with questions that go beyond simply ‘who is here’ to ask instead about individual characteristics, behavior, and species interactions. Being able to efficiently and accurately uncover these more complex phenomena in biodiversity image data will be critical to fundamental science and real-world impacts in ecology and conservation.”

Vendrow, Pantazis, and Beery wrote the paper with iNaturalist software engineer Alexander Shepard, University College London professors Gabriel Brostow and Kate Jones, University of Edinburgh associate professor and co-senior author Oisin Mac Aodha, and University of Massachusetts at Amherst Assistant Professor Grant Van Horn, who served as co-senior author. Their work was supported, in part, by the Generative AI Laboratory at the University of Edinburgh, the U.S. National Science Foundation/Natural Sciences and Engineering Research Council of Canada Global Center on AI and Biodiversity Change, a Royal Society Research Grant, and the Biome Health Project funded by the World Wildlife Fund United Kingdom.

Monitoring electrical signals in biological systems helps scientists understand how cells communicate, which can aid in the diagnosis and treatment of conditions like arrhythmia and Alzheimer’s.

But devices that record electrical signals in cell cultures and other liquid environments often use wires to connect each electrode on the device to its respective amplifier. Because only so many wires can be connected to the device, this restricts the number of recording sites, limiting the information that can be collected from cells.

MIT researchers have now developed a biosensing technique that eliminates the need for wires. Instead, tiny, wireless antennas use light to detect minute electrical signals.

Small electrical changes in the surrounding liquid environment alter how the antennas scatter the light. Using an array of tiny antennas, each of which is one-hundredth the width of a human hair, the researchers could measure electrical signals exchanged between cells, with extreme spatial resolution.

The devices, which are durable enough to continuously record signals for more than 10 hours, could help biologists understand how cells communicate in response to changes in their environment. In the long run, such scientific insights could pave the way for advancements in diagnosis, spur the development of targeted treatments, and enable more precision in the evaluation of new therapies.

“Being able to record the electrical activity of cells with high throughput and high resolution remains a real problem. We need to try some innovative ideas and alternate approaches,” says Benoît Desbiolles, a former postdoc in the MIT Media Lab and lead author of a paper on the devices.

He is joined on the paper by Jad Hanna, a visiting student in the Media Lab; former visiting student Raphael Ausilio; former postdoc Marta J. I. Airaghi Leccardi; Yang Yu, a scientist at Raith America, Inc.; and senior author Deblina Sarkar, the AT&T Career Development Assistant Professor in the Media Lab and MIT Center for Neurobiological Engineering and head of the Nano-Cybernetic Biotrek Lab. The research appears today in Science Advances.

“Bioelectricity is fundamental to the functioning of cells and different life processes. However, recording such electrical signals precisely has been challenging,” says Sarkar. “The organic electro-scattering antennas (OCEANs) we developed enable recording of electrical signals wirelessly with micrometer spatial resolution from thousands of recording sites simultaneously. This can create unprecedented opportunities for understanding fundamental biology and altered signaling in diseased states as well as for screening the effect of different therapeutics to enable novel treatments.”

Biosensing with light

The researchers set out to design a biosensing device that didn’t need wires or amplifiers. Such a device would be easier to use for biologists who may not be familiar with electronic instruments.

“We wondered if we could make a device that converts the electrical signals to light and then use an optical microscope, the kind that is available in every biology lab, to probe these signals,” Desbiolles says.

Initially, they used a special polymer called PEDOT:PSS to design nanoscale transducers that incorporated tiny pieces of gold filament. Gold nanoparticles were supposed to scatter the light — a process that would be induced and modulated by the polymer. But the results weren’t matching up with their theoretical model.

The researchers tried removing the gold and, surprisingly, the results matched the model much more closely.

“It turns out we weren’t measuring signals from the gold, but from the polymer itself. This was a very surprising but exciting result. We built on that finding to develop organic electro-scattering antennas,” he says.

The organic electro-scattering antennas, or OCEANs, are composed of PEDOT:PSS. This polymer attracts or repulses positive ions from the surrounding liquid environment when there is electrical activity nearby. This modifies its chemical configuration and electronic structure, altering an optical property known as its refractive index, which changes how it scatters light.

When researchers shine light onto the antenna, the intensity of the light changes in proportion to the electrical signal present in the liquid.

With thousands or even millions of tiny antennas in an array, each only 1 micrometer wide, the researchers can capture the scattered light with an optical microscope and measure electrical signals from cells with high resolution. Because each antenna is an independent sensor, the researchers do not need to pool the contribution of multiple antennas to monitor electrical signals, which is why OCEANs can detect signals with micrometer resolution.

Intended for in vitro studies, OCEAN arrays are designed to have cells cultured directly on top of them and put under an optical microscope for analysis.

“Growing” antennas on a chip

Key to the devices is the precision with which the researchers can fabricate arrays in the MIT.nano facilities.

They start with a glass substrate and deposit layers of conductive then insulating material on top, each of which is optically transparent. Then they use a focused ion beam to cut hundreds of nanoscale holes into the top layers of the device. This special type of focused ion beam enables high-throughput nanofabrication.

“This instrument is basically like a pen where you can etch anything with a 10-nanometer resolution,” he says.

They submerge the chip in a solution that contains the precursor building blocks for the polymer. By applying an electric current to the solution, that precursor material is attracted into the tiny holes on the chip, and mushroom-shaped antennas “grow” from the bottom up.

The entire fabrication process is relatively fast, and the researchers could use this technique to make a chip with millions of antennas.

“This technique could be easily adapted so it is fully scalable. The limiting factor is how many antennas we can image at the same time,” he says.

The researchers optimized the dimensions of the antennas and adjusted parameters, which enabled them to achieve high enough sensitivity to monitor signals with voltages as low as 2.5 millivolts in simulated experiments. Signals sent by neurons for communication are usually around 100 millivolts.

“Because we took the time to really dig in and understand the theoretical model behind this process, we can maximize the sensitivity of the antennas,” he says.

OCEANs also responded to changing signals in only a few milliseconds, enabling them to record electrical signals with fast kinetics. Moving forward, the researchers want to test the devices with real cell cultures. They also want to reshape the antennas so they can penetrate cell membranes, enabling more precise signal detection.

In addition, they want to study how OCEANs could be integrated into nanophotonic devices, which manipulate light at the nanoscale for next-generation sensors and optical devices.

This research is funded, in part, by the U.S. National Institutes of Health and the Swiss National Science Foundation. Research reported in this press release was supported by the National Heart, Lung, and Blood Institute (NHLBI) of the National Institutes of Health and does not necessarily represent the official views of the NIH.

Crafting a unique and promising research hypothesis is a fundamental skill for any scientist. It can also be time consuming: New PhD candidates might spend the first year of their program trying to decide exactly what to explore in their experiments. What if artificial intelligence could help?

MIT researchers have created a way to autonomously generate and evaluate promising research hypotheses across fields, through human-AI collaboration. In a new paper, they describe how they used this framework to create evidence-driven hypotheses that align with unmet research needs in the field of biologically inspired materials.

Published Wednesday in Advanced Materials, the study was co-authored by Alireza Ghafarollahi, a postdoc in the Laboratory for Atomistic and Molecular Mechanics (LAMM), and Markus Buehler, the Jerry McAfee Professor in Engineering in MIT’s departments of Civil and Environmental Engineering and of Mechanical Engineering and director of LAMM.

The framework, which the researchers call SciAgents, consists of multiple AI agents, each with specific capabilities and access to data, that leverage “graph reasoning” methods, where AI models utilize a knowledge graph that organizes and defines relationships between diverse scientific concepts. The multi-agent approach mimics the way biological systems organize themselves as groups of elementary building blocks. Buehler notes that this “divide and conquer” principle is a prominent paradigm in biology at many levels, from materials to swarms of insects to civilizations — all examples where the total intelligence is much greater than the sum of individuals’ abilities.

“By using multiple AI agents, we’re trying to simulate the process by which communities of scientists make discoveries,” says Buehler. “At MIT, we do that by having a bunch of people with different backgrounds working together and bumping into each other at coffee shops or in MIT’s Infinite Corridor. But that's very coincidental and slow. Our quest is to simulate the process of discovery by exploring whether AI systems can be creative and make discoveries.”

Automating good ideas

As recent developments have demonstrated, large language models (LLMs) have shown an impressive ability to answer questions, summarize information, and execute simple tasks. But they are quite limited when it comes to generating new ideas from scratch. The MIT researchers wanted to design a system that enabled AI models to perform a more sophisticated, multistep process that goes beyond recalling information learned during training, to extrapolate and create new knowledge.

The foundation of their approach is an ontological knowledge graph, which organizes and makes connections between diverse scientific concepts. To make the graphs, the researchers feed a set of scientific papers into a generative AI model. In previous work, Buehler used a field of math known as category theory to help the AI model develop abstractions of scientific concepts as graphs, rooted in defining relationships between components, in a way that could be analyzed by other models through a process called graph reasoning. This focuses AI models on developing a more principled way to understand concepts; it also allows them to generalize better across domains.

“This is really important for us to create science-focused AI models, as scientific theories are typically rooted in generalizable principles rather than just knowledge recall,” Buehler says. “By focusing AI models on ‘thinking’ in such a manner, we can leapfrog beyond conventional methods and explore more creative uses of AI.”

For the most recent paper, the researchers used about 1,000 scientific studies on biological materials, but Buehler says the knowledge graphs could be generated using far more or fewer research papers from any field.

With the graph established, the researchers developed an AI system for scientific discovery, with multiple models specialized to play specific roles in the system. Most of the components were built off of OpenAI’s ChatGPT-4 series models and made use of a technique known as in-context learning, in which prompts provide contextual information about the model’s role in the system while allowing it to learn from data provided.

The individual agents in the framework interact with each other to collectively solve a complex problem that none of them would be able to do alone. The first task they are given is to generate the research hypothesis. The LLM interactions start after a subgraph has been defined from the knowledge graph, which can happen randomly or by manually entering a pair of keywords discussed in the papers.

In the framework, a language model the researchers named the “Ontologist” is tasked with defining scientific terms in the papers and examining the connections between them, fleshing out the knowledge graph. A model named “Scientist 1” then crafts a research proposal based on factors like its ability to uncover unexpected properties and novelty. The proposal includes a discussion of potential findings, the impact of the research, and a guess at the underlying mechanisms of action. A “Scientist 2” model expands on the idea, suggesting specific experimental and simulation approaches and making other improvements. Finally, a “Critic” model highlights its strengths and weaknesses and suggests further improvements.

“It’s about building a team of experts that are not all thinking the same way,” Buehler says. “They have to think differently and have different capabilities. The Critic agent is deliberately programmed to critique the others, so you don't have everybody agreeing and saying it’s a great idea. You have an agent saying, ‘There’s a weakness here, can you explain it better?’ That makes the output much different from single models.”

Other agents in the system are able to search existing literature, which provides the system with a way to not only assess feasibility but also create and assess the novelty of each idea.

Making the system stronger

To validate their approach, Buehler and Ghafarollahi built a knowledge graph based on the words “silk” and “energy intensive.” Using the framework, the “Scientist 1” model proposed integrating silk with dandelion-based pigments to create biomaterials with enhanced optical and mechanical properties. The model predicted the material would be significantly stronger than traditional silk materials and require less energy to process.

Scientist 2 then made suggestions, such as using specific molecular dynamic simulation tools to explore how the proposed materials would interact, adding that a good application for the material would be a bioinspired adhesive. The Critic model then highlighted several strengths of the proposed material and areas for improvement, such as its scalability, long-term stability, and the environmental impacts of solvent use. To address those concerns, the Critic suggested conducting pilot studies for process validation and performing rigorous analyses of material durability.

The researchers also conducted other experiments with randomly chosen keywords, which produced various original hypotheses about more efficient biomimetic microfluidic chips, enhancing the mechanical properties of collagen-based scaffolds, and the interaction between graphene and amyloid fibrils to create bioelectronic devices.

“The system was able to come up with these new, rigorous ideas based on the path from the knowledge graph,” Ghafarollahi says. “In terms of novelty and applicability, the materials seemed robust and novel. In future work, we’re going to generate thousands, or tens of thousands, of new research ideas, and then we can categorize them, try to understand better how these materials are generated and how they could be improved further.”

Going forward, the researchers hope to incorporate new tools for retrieving information and running simulations into their frameworks. They can also easily swap out the foundation models in their frameworks for more advanced models, allowing the system to adapt with the latest innovations in AI.

“Because of the way these agents interact, an improvement in one model, even if it’s slight, has a huge impact on the overall behaviors and output of the system,” Buehler says.

Since releasing a preprint with open-source details of their approach, the researchers have been contacted by hundreds of people interested in using the frameworks in diverse scientific fields and even areas like finance and cybersecurity.

“There’s a lot of stuff you can do without having to go to the lab,” Buehler says. “You want to basically go to the lab at the very end of the process. The lab is expensive and takes a long time, so you want a system that can drill very deep into the best ideas, formulating the best hypotheses and accurately predicting emergent behaviors. Our vision is to make this easy to use, so you can use an app to bring in other ideas or drag in datasets to really challenge the model to make new discoveries.”

On June 18, 2023, the Titan submersible was about an hour-and-a-half into its two-hour descent to the Titanic wreckage at the bottom of the Atlantic Ocean when it lost contact with its support ship. This cease in communication set off a frantic search for the tourist submersible and five passengers onboard, located about two miles below the ocean's surface.

Deep-ocean search and recovery is one of the many missions of military services like the U.S. Coast Guard Office of Search and Rescue and the U.S. Navy Supervisor of Salvage and Diving. For this mission, the longest delays come from transporting search-and-rescue equipment via ship to the area of interest and comprehensively surveying that area. A search operation on the scale of that for Titan — which was conducted 420 nautical miles from the nearest port and covered 13,000 square kilometers, an area roughly twice the size of Connecticut — could take weeks to complete. The search area for Titan is considered relatively small, focused on the immediate vicinity of the Titanic. When the area is less known, operations could take months. (A remotely operated underwater vehicle deployed by a Canadian vessel ended up finding the debris field of Titan on the seafloor, four days after the submersible had gone missing.)

A research team from MIT Lincoln Laboratory and the MIT Department of Mechanical Engineering's Ocean Science and Engineering lab is developing a surface-based sonar system that could accelerate the timeline for small- and large-scale search operations to days. Called the Autonomous Sparse-Aperture Multibeam Echo Sounder, the system scans at surface-ship rates while providing sufficient resolution to find objects and features in the deep ocean, without the time and expense of deploying underwater vehicles. The echo sounder — which features a large sonar array using a small set of autonomous surface vehicles (ASVs) that can be deployed via aircraft into the ocean — holds the potential to map the seabed at 50 times the coverage rate of an underwater vehicle and 100 times the resolution of a surface vessel.

"Our array provides the best of both worlds: the high resolution of underwater vehicles and the high coverage rate of surface ships," says co–principal investigator Andrew March, assistant leader of the laboratory's Advanced Undersea Systems and Technology Group. "Though large surface-based sonar systems at low frequency have the potential to determine the materials and profiles of the seabed, they typically do so at the expense of resolution, particularly with increasing ocean depth. Our array can likely determine this information, too, but at significantly enhanced resolution in the deep ocean."

Underwater unknown

Oceans cover 71 percent of Earth's surface, yet more than 80 percent of this underwater realm remains undiscovered and unexplored. Humans know more about the surface of other planets and the moon than the bottom of our oceans. High-resolution seabed maps would not only be useful to find missing objects like ships or aircraft, but also to support a host of other scientific applications: understanding Earth's geology, improving forecasting of ocean currents and corresponding weather and climate impacts, uncovering archaeological sites, monitoring marine ecosystems and habitats, and identifying locations containing natural resources such as mineral and oil deposits.

Scientists and governments worldwide recognize the importance of creating a high-resolution global map of the seafloor; the problem is that no existing technology can achieve meter-scale resolution from the ocean surface. The average depth of our oceans is approximately 3,700 meters. However, today's technologies capable of finding human-made objects on the seabed or identifying person-sized natural features — these technologies include sonar, lidar, cameras, and gravitational field mapping — have a maximum range of less than 1,000 meters through water.

Ships with large sonar arrays mounted on their hull map the deep ocean by emitting low-frequency sound waves that bounce off the seafloor and return as echoes to the surface. Operation at low frequencies is necessary because water readily absorbs high-frequency sound waves, especially with increasing depth; however, such operation yields low-resolution images, with each image pixel representing a football field in size. Resolution is also restricted because sonar arrays installed on large mapping ships are already using all of the available hull space, thereby capping the sonar beam's aperture size. By contrast, sonars on autonomous underwater vehicles (AUVs) that operate at higher frequencies within a few hundred meters of the seafloor generate maps with each pixel representing one square meter or less, resulting in 10,000 times more pixels in that same football field–sized area. However, this higher resolution comes with trade-offs: AUVs are time-consuming and expensive to deploy in the deep ocean, limiting the amount of seafloor that can be mapped; they have a maximum range of about 1,000 meters before their high-frequency sound gets absorbed; and they move at slow speeds to conserve power. The area-coverage rate of AUVs performing high-resolution mapping is about 8 square kilometers per hour; surface vessels map the deep ocean at more than 50 times that rate.

A solution surfaces

The Autonomous Sparse-Aperture Multibeam Echo Sounder could offer a cost-effective approach to high-resolution, rapid mapping of the deep seafloor from the ocean's surface. A collaborative fleet of about 20 ASVs, each hosting a small sonar array, effectively forms a single sonar array 100 times the size of a large sonar array installed on a ship. The large aperture achieved by the array (hundreds of meters) produces a narrow beam, which enables sound to be precisely steered to generate high-resolution maps at low frequency. Because very few sonars are installed relative to the array's overall size (i.e., a sparse aperture), the cost is tractable.

However, this collaborative and sparse setup introduces some operational challenges. First, for coherent 3D imaging, the relative position of each ASV's sonar subarray must be accurately tracked through dynamic ocean-induced motions. Second, because sonar elements are not placed directly next to each other without any gaps, the array suffers from a lower signal-to-noise ratio and is less able to reject noise coming from unintended or undesired directions. To mitigate these challenges, the team has been developing a low-cost precision-relative navigation system and leveraging acoustic signal processing tools and new ocean-field estimation algorithms. The MIT campus collaborators are developing algorithms for data processing and image formation, especially to estimate depth-integrated water-column parameters. These enabling technologies will help account for complex ocean physics, spanning physical properties like temperature, dynamic processes like currents and waves, and acoustic propagation factors like sound speed.

Processing for all required control and calculations could be completed either remotely or onboard the ASVs. For example, ASVs deployed from a ship or flying boat could be controlled and guided remotely from land via a satellite link or from a nearby support ship (with direct communications or a satellite link), and left to map the seabed for weeks or months at a time until maintenance is needed. Sonar-return health checks and coarse seabed mapping would be conducted on board, while full, high-resolution reconstruction of the seabed would require a supercomputing infrastructure on land or on a support ship.

"Deploying vehicles in an area and letting them map for extended periods of time without the need for a ship to return home to replenish supplies and rotate crews would significantly simplify logistics and operating costs," says co–principal investigator Paul Ryu, a researcher in the Advanced Undersea Systems and Technology Group.

Since beginning their research in 2018, the team has turned their concept into a prototype. Initially, the scientists built a scale model of a sparse-aperture sonar array and tested it in a water tank at the laboratory's Autonomous Systems Development Facility. Then, they prototyped an ASV-sized sonar subarray and demonstrated its functionality in Gloucester, Massachusetts. In follow-on sea tests in Boston Harbor, they deployed an 8-meter array containing multiple subarrays equivalent to 25 ASVs locked together; with this array, they generated 3D reconstructions of the seafloor and a shipwreck. Most recently, the team fabricated, in collaboration with Woods Hole Oceanographic Institution, a first-generation, 12-foot-long, all-electric ASV prototype carrying a sonar array underneath. With this prototype, they conducted preliminary relative navigation testing in Woods Hole, Massachusetts and Newport, Rhode Island. Their full deep-ocean concept calls for approximately 20 such ASVs of a similar size, likely powered by wave or solar energy.

This work was funded through Lincoln Laboratory's internally administered R&D portfolio on autonomous systems. The team is now seeking external sponsorship to continue development of their ocean floor–mapping technology, which was recognized with a 2024 R&D 100 Award. 

From studies of the connections between neurons to interactions between the nervous and immune systems to the complex ways in which people understand not just language, but also the unspoken nuances of conversation, new research projects at MIT supported by the Simons Center for the Social Brain are bringing a rich diversity of perspectives to advancing the field’s understanding of autism.

As six speakers lined up to describe their projects at a Simons Center symposium Nov. 15, MIT School of Science dean Nergis Mavalvala articulated what they were all striving for: “Ultimately, we want to seek understanding — not just the type that tells us how physiological differences in the inner workings of the brain produce differences in behavior and cognition, but also the kind of understanding that improves inclusion and quality of life for people living with autism spectrum disorders.”

Simons Center director Mriganka Sur, Newton Professor of Neuroscience in The Picower Institute for Learning and Memory and Department of Brain and Cognitive Sciences (BCS), said that even though the field still lacks mechanism-based treatments or reliable biomarkers for autism spectrum disorders, he is optimistic about the discoveries and new research MIT has been able to contribute. MIT research has led to five clinical trials so far, and he praised the potential for future discovery, for instance in the projects showcased at the symposium.

“We are, I believe, at a frontier — at a moment where a lot of basic science is coming together with the vision that we could use that science for the betterment of people,” Sur said.

The Simons Center funds that basic science research in two main ways that each encourage collaboration, Sur said: large-scale projects led by faculty members across several labs, and fellowships for postdocs who are mentored by two faculty members, thereby bringing together two labs. The symposium featured talks and panel discussions by faculty and fellows leading new research.

In a different vein, Associate Professor Ev Fedorenko of The McGovern Institute for Brain Research and BCS is leading a seven-lab collaboration aimed at understanding the cognitive and neural infrastructure that enables people to engage in conversation, which involves not only the language spoken but also facial expressions, tone of voice, and social context. Critical to this effort, she said, is going beyond previous work that studied each related brain area in isolation to understand the capability as a unified whole. A key insight, she said, is that they are all nearby each other in the lateral temporal cortex.

“Going beyond these individual components we can start asking big questions like, what are the broad organizing principles of this part of the brain?,” Fedorenko said. “Why does it have this particular arrangement of areas, and how do these work together to exchange information to create the unified percept of another individual we’re interacting with?”

While Choi and Fedorenko are looking at factors that account for differences in social behavior in autism, Picower Professor Earl K. Miller of The Picower Institute and BCS is leading a project that focuses on another phenomenon: the feeling of sensory overload that many autistic people experience. Research in Miller’s lab has shown that the brain’s ability to make predictions about sensory stimuli, which is critical to filtering out mundane signals so attention can be focused on new ones, depends on a cortex-wide coordination of the activity of millions of neurons implemented by high frequency “gamma” brain waves and lower-frequency “beta” waves. Working with animal models and human volunteers at Boston Children’s Hospital (BCH), Miller said his team is testing the idea that there may be a key difference in these brain wave dynamics in the autistic brain that could be addressed with closed-loop brain wave stimulation technology.

Simons postdoc Lukas Vogelsang, who is based in BCS Professor Pawan Sinha’s lab, is looking at potential differences in prediction between autistic and non-autistic individuals in a different way: through experiments with volunteers that aim to tease out how these differences are manifest in behavior. For instance, he’s finding that in at least one prediction task that requires participants to discern the probability of an event from provided cues, autistic people exhibit lower performance levels and undervalue the predictive significance of the cues, while non-autistic people slightly overvalue it. Vogelsang is co-advised by BCH researcher and Harvard Medical School Professor Charles Nelson.

Fundamentally, the broad-scale behaviors that emerge from coordinated brain-wide neural activity begins with the molecular details of how neurons connect with each other at circuit junctions called synapses. In her research based in The Picower Institute lab of Menicon Professor Troy Littleton, Simons postdoc Chhavi Sood is using the genetically manipulable model of the fruit fly to investigate how mutations in the autism-associated protein FMRP may alter the expression of molecular gates regulating ion exchange at the synapse , which would in turn affect how frequently and strongly a pre-synaptic neuron excites a post-synaptic one. The differences she is investigating may be a molecular mechanism underlying neural hyperexcitability in fragile X syndrome, a profound autism spectrum disorder.

In her talk, Simons postdoc Lace Riggs, based in The McGovern Institute lab of Poitras Professor of Neuroscience Guoping Feng, emphasized how many autism-associated mutations in synaptic proteins promote pathological anxiety. She described her research that is aimed at discerning where in the brain’s neural circuitry that vulnerability might lie. In her ongoing work, Riggs is zeroing in on a novel thalamocortical circuit between the anteromedial nucleus of the thalamus and the cingulate cortex, which she found drives anxiogenic states. Riggs is co-supervised by Professor Fan Wang.

After the wide-ranging talks, supplemented by further discussion at the panels, the last word came via video conference from Kelsey Martin, executive vice president of the Simons Foundation Autism Research Initiative. Martin emphasized that fundamental research, like that done at the Simons Center, is the key to developing future therapies and other means of supporting members of the autism community.

“We believe so strongly that understanding the basic mechanisms of autism is critical to being able to develop translational and clinical approaches that are going to impact the lives of autistic individuals and their families,” she said.

From studies of synapses to circuits to behavior, MIT researchers and their collaborators are striving for exactly that impact.

The electronics industry is approaching a limit to the number of transistors that can be packed onto the surface of a computer chip. So, chip manufacturers are looking to build up rather than out.

Instead of squeezing ever-smaller transistors onto a single surface, the industry is aiming to stack multiple surfaces of transistors and semiconducting elements — akin to turning a ranch house into a high-rise. Such multilayered chips could handle exponentially more data and carry out many more complex functions than today’s electronics.

A significant hurdle, however, is the platform on which chips are built. Today, bulky silicon wafers serve as the main scaffold on which high-quality, single-crystalline semiconducting elements are grown. Any stackable chip would have to include thick silicon “flooring” as part of each layer, slowing down any communication between functional semiconducting layers.

Now, MIT engineers have found a way around this hurdle, with a multilayered chip design that doesn’t require any silicon wafer substrates and works at temperatures low enough to preserve the underlying layer’s circuitry.

In a study appearing today in the journal Nature, the team reports using the new method to fabricate a multilayered chip with alternating layers of high-quality semiconducting material grown directly on top of each other.

The method enables engineers to build high-performance transistors and memory and logic elements on any random crystalline surface — not just on the bulky crystal scaffold of silicon wafers. Without these thick silicon substrates, multiple semiconducting layers can be in more direct contact, leading to better and faster communication and computation between layers, the researchers say.

The researchers envision that the method could be used to build AI hardware, in the form of stacked chips for laptops or wearable devices, that would be as fast and powerful as today’s supercomputers and could store huge amounts of data on par with physical data centers.

“This breakthrough opens up enormous potential for the semiconductor industry, allowing chips to be stacked without traditional limitations,” says study author Jeehwan Kim, associate professor of mechanical engineering at MIT. “This could lead to orders-of-magnitude improvements in computing power for applications in AI, logic, and memory.”

The study’s MIT co-authors include first author Ki Seok Kim, Seunghwan Seo, Doyoon Lee, Jung-El Ryu, Jekyung Kim, Jun Min Suh, June-chul Shin, Min-Kyu Song, Jin Feng, and Sangho Lee, along with collaborators from Samsung Advanced Institute of Technology, Sungkyunkwan University in South Korea, and the University of Texas at Dallas.

Seed pockets

In 2023, Kim’s group reported that they developed a method to grow high-quality semiconducting materials on amorphous surfaces, similar to the diverse topography of semiconducting circuitry on finished chips. The material that they grew was a type of 2D material known as transition-metal dichalcogenides, or TMDs, considered a promising successor to silicon for fabricating smaller, high-performance transistors. Such 2D materials can maintain their semiconducting properties even at scales as small as a single atom, whereas silicon’s performance sharply degrades.

In their previous work, the team grew TMDs on silicon wafers with amorphous coatings, as well as over existing TMDs. To encourage atoms to arrange themselves into high-quality single-crystalline form, rather than in random, polycrystalline disorder, Kim and his colleagues first covered a silicon wafer in a very thin film, or “mask” of silicon dioxide, which they patterned with tiny openings, or pockets. They then flowed a gas of atoms over the mask and found that atoms settled into the pockets as “seeds.” The pockets confined the seeds to grow in regular, single-crystalline patterns.

But at the time, the method only worked at around 900 degrees Celsius.

“You have to grow this single-crystalline material below 400 Celsius, otherwise the underlying circuitry is completely cooked and ruined,” Kim says. “So, our homework was, we had to do a similar technique at temperatures lower than 400 Celsius. If we could do that, the impact would be substantial.”

Building up

In their new work, Kim and his colleagues looked to fine-tune their method in order to grow single-crystalline 2D materials at temperatures low enough to preserve any underlying circuitry. They found a surprisingly simple solution in metallurgy — the science and craft of metal production. When metallurgists pour molten metal into a mold, the liquid slowly “nucleates,” or forms grains that grow and merge into a regularly patterned crystal that hardens into solid form. Metallurgists have found that this nucleation occurs most readily at the edges of a mold into which liquid metal is poured.

“It’s known that nucleating at the edges requires less energy — and heat,” Kim says. “So we borrowed this concept from metallurgy to utilize for future AI hardware.”

The team looked to grow single-crystalline TMDs on a silicon wafer that already has been fabricated with transistor circuitry. They first covered the circuitry with a mask of silicon dioxide, just as in their previous work. They then deposited “seeds” of TMD at the edges of each of the mask’s pockets and found that these edge seeds grew into single-crystalline material at temperatures as low as 380 degrees Celsius, compared to seeds that started growing in the center, away from the edges of each pocket, which required higher temperatures to form single-crystalline material.

Going a step further, the researchers used the new method to fabricate a multilayered chip with alternating layers of two different TMDs — molybdenum disulfide, a promising material candidate for fabricating n-type transistors; and tungsten diselenide, a material that has potential for being made into p-type transistors. Both p- and n-type transistors are the electronic building blocks for carrying out any logic operation. The team was able to grow both materials in single-crystalline form, directly on top of each other, without requiring any intermediate silicon wafers. Kim says the method will effectively double the density of a chip’s semiconducting elements, and particularly, metal-oxide semiconductor (CMOS), which is a basic building block of a modern logic circuitry.

“A product realized by our technique is not only a 3D logic chip but also 3D memory and their combinations,” Kim says. “With our growth-based monolithic 3D method, you could grow tens to hundreds of logic and memory layers, right on top of each other, and they would be able to communicate very well.”

“Conventional 3D chips have been fabricated with silicon wafers in-between, by drilling holes through the wafer — a process which limits the number of stacked layers, vertical alignment resolution, and yields,” first author Kiseok Kim adds. “Our growth-based method addresses all of those issues at once.” 

To commercialize their stackable chip design further, Kim has recently spun off a company, FS2 (Future Semiconductor 2D materials).

“We so far show a concept at a small-scale device arrays,” he says. “The next step is scaling up to show professional AI chip operation.”

This research is supported, in part, by Samsung Advanced Institute of Technology and the U.S. Air Force Office of Scientific Research. 

MIT physicists have created a new and long-lasting magnetic state in a material, using only light.

In a study appearing today in Nature, the researchers report using a terahertz laser — a light source that oscillates more than a trillion times per second — to directly stimulate atoms in an antiferromagnetic material. The laser’s oscillations are tuned to the natural vibrations among the material’s atoms, in a way that shifts the balance of atomic spins toward a new magnetic state.

The results provide a new way to control and switch antiferromagnetic materials, which are of interest for their potential to advance information processing and memory chip technology.

In common magnets, known as ferromagnets, the spins of atoms point in the same direction, in a way that the whole can be easily influenced and pulled in the direction of any external magnetic field. In contrast, antiferromagnets are composed of atoms with alternating spins, each pointing in the opposite direction from its neighbor. This up, down, up, down order essentially cancels the spins out, giving antiferromagnets a net zero magnetization that is impervious to any magnetic pull.

If a memory chip could be made from antiferromagnetic material, data could be “written” into microscopic regions of the material, called domains. A certain configuration of spin orientations (for example, up-down) in a given domain would represent the classical bit “0,” and a different configuration (down-up) would mean “1.” Data written on such a chip would be robust against outside magnetic influence.

For this and other reasons, scientists believe antiferromagnetic materials could be a more robust alternative to existing magnetic-based storage technologies. A major hurdle, however, has been in how to control antiferromagnets in a way that reliably switches the material from one magnetic state to another.

“Antiferromagnetic materials are robust and not influenced by unwanted stray magnetic fields,” says Nuh Gedik, the Donner Professor of Physics at MIT. “However, this robustness is a double-edged sword; their insensitivity to weak magnetic fields makes these materials difficult to control.”

Using carefully tuned terahertz light, the MIT team was able to controllably switch an antiferromagnet to a new magnetic state. Antiferromagnets could be incorporated into future memory chips that store and process more data while using less energy and taking up a fraction of the space of existing devices, owing to the stability of magnetic domains.

“Generally, such antiferromagnetic materials are not easy to control,” Gedik says. “Now we have some knobs to be able to tune and tweak them.”

Gedik is the senior author of the new study, which also includes MIT co-authors Batyr Ilyas, Tianchuang Luo, Alexander von Hoegen, Zhuquan Zhang, and Keith Nelson, along with collaborators at the Max Planck Institute for the Structure and Dynamics of Matter in Germany, University of the Basque Country in Spain, Seoul National University, and the Flatiron Institute in New York.

Off balance

Gedik’s group at MIT develops techniques to manipulate quantum materials in which interactions among atoms can give rise to exotic phenomena.

“In general, we excite materials with light to learn more about what holds them together fundamentally,” Gedik says. “For instance, why is this material an antiferromagnet, and is there a way to perturb microscopic interactions such that it turns into a ferromagnet?”

In their new study, the team worked with FePS₃ — a material that transitions to an antiferromagnetic phase at a critical temperature of around 118 kelvins (-247 degrees Fahrenheit).

The team suspected they might control the material’s transition by tuning into its atomic vibrations.

“In any solid, you can picture it as different atoms that are periodically arranged, and between atoms are tiny springs,” von Hoegen explains. “If you were to pull one atom, it would vibrate at a characteristic frequency which typically occurs in the terahertz range.”

The way in which atoms vibrate also relates to how their spins interact with each other. The team reasoned that if they could stimulate the atoms with a terahertz source that oscillates at the same frequency as the atoms’ collective vibrations, called phonons, the effect could also nudge the atoms’ spins out of their perfectly balanced, magnetically alternating alignment. Once knocked out of balance, atoms should have larger spins in one direction than the other, creating a preferred orientation that would shift the inherently nonmagnetized material into a new magnetic state with finite magnetization.

“The idea is that you can kill two birds with one stone: You excite the atoms’ terahertz vibrations, which also couples to the spins,” Gedik says.

Shake and write

To test this idea, the team worked with a sample of FePS₃ that was synthesized by colleages at Seoul National University. They placed the sample in a vacuum chamber and cooled it down to temperatures at and below 118 K. They then generated a terahertz pulse by aiming a beam of near-infrared light through an organic crystal, which transformed the light into the terahertz frequencies. They then directed this terahertz light toward the sample.

“This terahertz pulse is what we use to create a change in the sample,” Luo says. “It’s like ‘writing’ a new state into the sample.”

To confirm that the pulse triggered a change in the material’s magnetism, the team also aimed two near-infrared lasers at the sample, each with an opposite circular polarization. If the terahertz pulse had no effect, the researchers should see no difference in the intensity of the transmitted infrared lasers.

“Just seeing a difference tells us the material is no longer the original antiferromagnet, and that we are inducing a new magnetic state, by essentially using terahertz light to shake the atoms,” Ilyas says.

Over repeated experiments, the team observed that a terahertz pulse successfully switched the previously antiferromagnetic material to a new magnetic state — a transition that persisted for a surprisingly long time, over several milliseconds, even after the laser was turned off.

“People have seen these light-induced phase transitions before in other systems, but typically they live for very short times on the order of a picosecond, which is a trillionth of a second,” Gedik says.

In just a few milliseconds, scientists now might have a decent window of time during which they could probe the properties of the temporary new state before it settles back into its inherent antiferromagnetism. Then, they might be able to identify new knobs to tweak antiferromagnets and optimize their use in next-generation memory storage technologies.

This research was supported, in part, by the U.S. Department of Energy, Materials Science and Engineering Division, Office of Basic Energy Sciences, and the Gordon and Betty Moore Foundation. 

Researchers have developed a model that explains how humans adapt continuously during complex tasks, like walking, while remaining stable.

The findings were detailed in a recent paper published in the journal Nature Communications authored by Nidhi Seethapathi, an assistant professor in MIT’s Department of Brain and Cognitive Sciences; Barrett C. Clark, a robotics software engineer at Bright Minds Inc.; and Manoj Srinivasan, an associate professor in the Department of Mechanical and Aerospace Engineering at Ohio State University.

In episodic tasks, like reaching for an object, errors during one episode do not affect the next episode. In tasks like locomotion, errors can have a cascade of short-term and long-term consequences to stability unless they are controlled. This makes the challenge of adapting locomotion in a new environment  more complex.

"Much of our prior theoretical understanding of adaptation has been limited to episodic tasks, such as reaching for an object in a novel environment," Seethapathi says. "This new theoretical model captures adaptation phenomena in continuous long-horizon tasks in multiple locomotor settings."

To build the model, the researchers identified general principles of locomotor adaptation across a variety of task settings, and  developed a unified modular and hierarchical model of locomotor adaptation, with each component having its own unique mathematical structure.

The resulting model successfully encapsulates how humans adapt their walking in novel settings such as on a split-belt treadmill with each foot at a different speed, wearing asymmetric leg weights, and wearing  an exoskeleton. The authors report that the model successfully reproduced human locomotor adaptation phenomena across novel settings in 10 prior studies and correctly predicted the adaptation behavior observed in two new experiments conducted as part of the study.

The model has potential applications in sensorimotor learning, rehabilitation, and wearable robotics.

"Having a model that can predict how a person will adapt to a new environment has immense utility for engineering better rehabilitation paradigms and wearable robot control," Seethapathi says. "You can think of a wearable robot itself as a new environment for the person to move in, and our model can be used to predict how a person will adapt for different robot settings. Understanding such human-robot adaptation is currently an experimentally intensive process, and our model  could help speed up the process by narrowing the search space."