Kategorie: KI Use Cases

MIT Associate Professor Jacob Andreas of the Department of Electrical Engineering and Computer Science [EECS] and MIT Associate Professor Brett McGuire of the Department of Chemistry have been selected as the winners of the 2026 Harold E. Edgerton Faculty Achievement Award. Established in 1982 as a permanent tribute to Institute Professor Emeritus Harold E. Edgerton’s great and enduring support for younger faculty members, this award is given annually in recognition of exceptional distinction in teaching, research, and service.

“The Department of Chemistry is extremely delighted to see Brett recognized for science that has changed how we think about carbon in space,” says Class of 1942 Professor of Chemistry and Department Head Matthew D. Shoulders. “Brett’s lab combines laboratory spectroscopy, radio astronomy, and sophisticated signal-analysis methods to pull definitive molecular fingerprints out of extraordinarily faint data. His discovery of polycyclic aromatic hydrocarbons in the cold interstellar medium has opened a powerful new window on astrochemistry. Moreover, Brett is inventing the creative and unique tools that make discoveries like this possible.”

“Jacob Andreas represents the very best of MIT EECS” says Asu Ozdaglar, EECS department head. “He is an innovative researcher whose work combines computational and linguistically informed approaches to build foundations of language learning. He is an extraordinary educator who has brought these forefront ideas into our core classes in natural language processing and machine learning. His ability to bridge foundational theory with real-world impact, while also advancing the social and ethical dimensions of computing, makes him truly deserving of the Edgerton Faculty Achievement Award.”

Andreas joined the MIT faculty in July 2019, and is affiliated with the Computer Science and Artificial Intelligence Laboratory. His work is in natural language processing (NLP), and more broadly in AI. He aims to understand the computational foundations of language learning, and to build intelligent systems that can learn from human guidance. Among other honors, Andreas has received Samsung’s AI Researcher of the Year award, MIT’s Kolokotrones and Junior Bose teaching awards, a 2024 Sloan Research Fellow award, and paper awards at the National Accrediting Agency for Clinical Laboratory Sciences, the International Conference on Machine Learning, and the Association for Computational Linguistics.

Andreas received his BS from Columbia University, his MPhil from Cambridge University (where he studied as a Churchill scholar), and his PhD in natural language processing from the University of California at Berkeley. His work in natural language processing has taken on thorny problems in the capability gap between humans and computers. “The defining feature of human language use is our capacity for compositional generalization,” explains Antonio Torralba, Delta Electronics Professor and faculty head of Artificial Intelligence and Decision-Making in the Department of EECS. “Many of the core challenges in natural language processing is addressed by simply training larger and larger neural models, but this kind of compositional generalization remains a persistent difficulty, and without the ability to generalize compositionally, the deep learning toolkit will never be robust enough for the most challenging real-world NLP tasks. Jacob’s work on compositional modeling draws new connections between NLP and work in computer vision and physics aimed at modeling systems governed by symmetries and other algebraic structures and, using them, they have been able to build NLP models exhibiting a number of new, human-like language acquisition behaviors, including one-shot word learning, learning via mutual exclusivity constraints, and learning of grammatical rules in extremely low-resource settings.”

Within EECS, Andreas has developed multiple advanced courses in natural language processing, as well as new exercises designed to get students to grapple with important social and ethical considerations in machine learning deployment. “Jacob has taken a leading role in completely modernizing and extending our course offerings in natural language processing,” says award nominator Leslie Pack Kaelbling, Panasonic Professor in the Department of EECS. “He has led the development of a modern two-course sequence, which is a cornerstone of the new AI+D [artificial intelligence and decision-making] major, routinely enrolling several hundred students each semester. His command of the area is broad and deep, and his classes integrate classical structural understanding of language with the most modern learning-based approaches. He has put MIT EECS on the worldwide map as a place to study natural language at every level.”

Brett McGuire joined the MIT faculty in 2020 and was promoted to associate professor in 2025. His research operates at the intersection of physical chemistry, molecular spectroscopy, and observational astrophysics, where he seeks to uncover how the chemical building blocks of life evolve alongside and help shape the birth of stars and planets. A former Jansky Fellow and then Hubble Postdoctoral Fellow at the National Radio Astronomy Observatory, McGuire has a BS in chemistry from the University of Illinois and a PhD in physical chemistry from Caltech. His honors include a 2026 Sloan Fellowship, the Beckman Young Investigator Award, the Helen B. Warner Prize for Astronomy, and the MIT Award for Teaching with Digital Technology.

The faculty who nominated McGuire for this award praised his extraordinary public outreach, his immediate willingness to take on teaching class 5.111 (Principles of Chemical Science), a General Institute Requirement (GIR) course comprised of 150–500 students, and his service to both the MIT and astrochemical communities.

“Brett is at the very top of astrochemical scientists in his age group due to his discovery of fused carbon ring compounds in the cold region of the ISM [interstellar medium], an observation that provides a route for carbon incorporation in planets,” says Sylvia Ceyer, the John C. Sheehan Professor of Chemistry in her nomination statement. “His extensive involvement in service-oriented activities within the astrochemical/physical community is highly unusual for a junior scientist, and is testament to the value that the astronomical community places in his wisdom and judgement. His phenomenal organizational skills have made his contributions to graduate admission protocols and seminar administration at MIT the envy of the department. And most importantly, Brett is a superb teacher, who cares deeply about students’ understanding and success, not only in his course, but in their future endeavors.”

“As an assistant professor, Brett volunteered to teach 5.111, a large GIR course with 150–500 students, and has received some of the best teaching evaluations among all faculty who have led the subject,” says Mei Hong, the David A. Leighty Professor of Chemistry. “He has a natural talent in explaining abstract physical chemistry concepts in an engaging manner. His slides, which he prepared from scratch instead of modifying from previous years’ material from other professors, are clear, and … the combination of lucid explanation and humor has generated great enthusiasm and interest in chemistry among students.”

Subject evaluations from McGuire’s courses praised his humor, the clarity of his explanations, and his ability to transform a lecture into a “science show.” “I haven’t felt this sort of desire for the depth of understanding in a subject beyond just a straight grade [in some time],” says one student. “Brett definitely stimulated that love of learning for me.” 

“Brett is an outstanding faculty member who is dedicated to fostering student learning and success,” says Jennifer Weisman, assistant director of academic programs in chemistry. “He is thoughtful, caring, and goes above and beyond to help his colleagues, students, and staff.”

“I’m thrilled to be selected for the Edgerton Award this year,” says McGuire. “The award is nominally for teaching, research, and service; MIT and the chemistry department in particular have been an incredible place to learn and grow in all these areas. I’m incredibly grateful for the mentorship, enthusiasm, and support I have received from my colleagues, from my students both in the lab and in the classroom, and from the MIT community during my time here. I look forward to many more years of exciting discovery together with this one-of-a-kind community.”

Artificial intelligence is already proving it can accelerate drug development and improve our understanding of disease. But to turn AI into novel treatments we need to get the latest, most powerful models into the hands of scientists.

The problem is that most scientists aren’t machine-learning experts. Now the company OpenProtein.AI is helping scientists stay on the cutting edge of AI with a no-code platform that gives them access to powerful foundation models and a suite of tools for designing proteins, predicting protein structure and function, and training models.

The company, founded by Tristan Bepler PhD ’20 and former MIT associate professor Tim Lu PhD ’07, is already equipping researchers in pharmaceutical and biotech companies of all sizes with its tools, including internally developed foundation models for protein engineering. OpenProtein.AI also offers its platform to scientists in academia for free.

“It’s a really exciting time right now because these models can not only make protein engineering more efficient — which shortens development cycles for therapeutics and industrial uses — they can also enhance our ability to design new proteins with specific traits,” Bepler says. “We’re also thinking about applying these approaches to non-protein modalities. The big picture is we’re creating a language for describing biological systems.”

Advancing biology with AI

Bepler came to MIT in 2014 as part of the Computational and Systems Biology PhD Program, studying under Bonnie Berger, MIT’s Simons Professor of Applied Mathematics. It was there that he realized how little we understand about the molecules that make up the building blocks of biology.

“We hadn’t characterized biomolecules and proteins well enough to create good predictive models of what, say, a whole genome circuit will do, or how a protein interaction network will behave,” Bepler recalls. “It got me interested in understanding proteins at a more fine-grained level.”

Bepler began exploring ways to predict the chains of amino acids that make up proteins by analyzing evolutionary data. This was before Google released AlphaFold, a powerful prediction model for protein structure. The work led to one of the first generative AI models for understanding and designing proteins — what the team calls a protein language model.

“I was really excited about the classical framework of proteins and the relationships between their sequence, structure, and function. We don’t understand those links well,” Bepler says. “So how could we use these foundation models to skip the ‘structure’ component and go straight from sequence to function?”

After earning his PhD in 2020, Bepler entered Lu’s lab in MIT’s Department of Biological Engineering as a postdoc.

“This was around the time when the idea of integrating AI with biology was starting to pick up,” Lu recalls. “Tristan helped us build better computational models for biologic design. We also realized there’s a disconnect between the most cutting-edge tools available and the biologists, who would love to use these things but don’t know how to code. OpenProtein came from the idea of broadening access to these tools.”

Bepler had worked at the forefront of AI as part of his PhD. He knew the technology could help scientists accelerate their work.

“We started with the idea to build a general-purpose platform for doing machine learning-in-the-loop protein engineering,” Bepler says. “We wanted to build something that was user friendly because machine-learning ideas are kind of esoteric. They require implementation, GPUs, fine-tuning, designing libraries of sequences. Especially at that time, it was a lot for biologists to learn.”

OpenProtein’s platform, in contrast, features an intuitive web interface for biologists to upload data and conduct protein engineering work with machine learning. It features a range of open-source models, including PoET, OpenProtein’s flagship protein language model.

PoET, short for Protein Evolutionary Transformer, was trained on protein groups to generate sets of related proteins. Bepler and his collaborators showed it could generalize about evolutionary constraints on proteins and incorporate new information on protein sequences without retraining, allowing other researchers to add experimental data to improve the model.

“Researchers can use their own data to train models and optimize protein sequences, and then they can use our other tools to analyze those proteins,” Bepler says. “People are generating libraries of protein sequences in silico [on computers] and then running them through predictive models to get validation and structural predictors. It’s basically a no-code front-end, but we also have APIs for people who want to access it with code.”

The models help researchers design proteins faster, then decide which ones are promising enough for further lab testing. Researchers can also input proteins of interest, and the models can generate new ones with similar properties.

Since its founding, OpenProtein’s team has continued to add tools to its platform for researchers regardless of their lab size or resources.

“We’ve tried really hard to make the platform an open-ended toolbox,” Bepler says. “It has specific workflows, but it’s not tied specifically to one protein function or class of proteins. One of the great things about these models is they are very good at understanding proteins broadly. They learn about the whole space of possible proteins.”

Enabling the next generation of therapies

The large pharmaceutical company Boehringer Ingelheim began using OpenProtein’s platform in early 2025. Recently, the companies announced an expanded collaboration that will see OpenProtein’s platform and models embedded into Boehringer Ingelheim’s work as it engineers proteins to treat diseases like cancer and autoimmune or inflammatory conditions.

Last year, OpenProtein also released a new version of its protein language model, PoET-2, that outperforms much larger models while using a small fraction of the computing resources and experimental data.

“We really want to solve the question of how we describe proteins,” Bepler says. “What’s the meaningful, domain-specific language of protein constraints we use as we generate them? How can we bring in more evolutionary constraints? How can we describe an enzymatic reaction a protein carries out such that a model can generate sequences to do that reaction?”

Moving forward, the founders are hoping to make models that factor in the changing, interconnected nature of protein function.

“The area I am excited about is going beyond protein binding events to use these models to predict and design dynamic features, where the protein has to engage two, three, or four biological mechanisms at the same time, or change its function after binding,” says Lu, who currently serves in an advisory role for the company.

As progress in AI races forward, OpenProtein continues to see its mission as giving scientists the best tools to develop new treatments faster.

“As work gets more complex, with approaches incorporating things like protein logic and dynamic therapies, the existing experimental toolsets become limiting,” Lu says. “It’s really important to create open ecosystems around AI and biology. There’s a risk that AI resources could get so concentrated that the average researcher can’t use them. Open access is super important for the scientific field to make progress.”

The electricity to an island goes out. To find the break in the underwater power cable, a ship pulls up the entire line or deploys remotely operated vehicles (ROVs) to traverse the line. But what if an autonomous underwater vehicle (AUV) could map the line and pinpoint the location of the fault for a diver to fix?

Such underwater human-robot teaming is the focus of an MIT Lincoln Laboratory project funded through an internally administered R&D portfolio on autonomous systems and carried out by the Advanced Undersea Systems and Technology Group. The project seeks to leverage the respective strengths of humans and robots to optimize maritime missions for the U.S. military, including critical infrastructure inspection and repair, search and rescue, harbor entry, and countermine operations.

„Divers and AUVs generally don’t team at all underwater,“ says principal investigator Madeline Miller. „Underwater missions requiring humans typically do so because they involve some sort of manipulation a robot can’t do, like repairing infrastructure or deactivating a mine. Even ROVs are challenging to work with underwater in very skilled manipulation tasks because the manipulators themselves aren’t agile enough.“

Beyond their superior dexterity, humans excel at recognizing objects underwater. But humans working underwater can’t perform complex computations or move very quickly, especially if they are carrying heavy equipment; robots have an edge over humans in processing power, high-speed mobility, and endurance. To combine these strengths, Miller and her team are developing hardware and algorithms for underwater navigation and perception — two key capabilities for effective human-robot teaming.

As Miller explains, divers may only have a compass and fin-kick counts to guide them. With few landmarks and potentially murky conditions caused by a lack of light at depth or the presence of biological matter in the water column, they can easily become disoriented and lost. For robots to help divers navigate, they need to perceive their environment. However, in the presence of darkness and turbidity, optical sensors (cameras) cannot generate images, while acoustic sensors (sonar) generate images that lack color and only show the shapes and shadows of objects in the scene. The historical lack of large, labeled sonar image datasets has hindered training of underwater perception algorithms. Even if data were available, the dynamic ocean can obscure the true nature of objects, confusing artificial intelligence. For instance, a downed aircraft broken into multiple pieces, or a tire covered in an overgrowth of mussels, may no longer resemble an aircraft or tire, respectively.

„Ultimately, we want to devise solutions for navigation and perception in expeditionary environments,“ Miller says. „For the missions we’re thinking about, there is limited or no opportunity to map out the area in advance. For the harbor entry mission, maybe you have a satellite map but no underwater map, for example.“

On the navigation side, Miller’s team picked up on work started by the MIT Marine Robotics Group, led by John Leonard, to develop diver-AUV teaming algorithms. With their navigation algorithms, Leonard’s group ran simulations under optimal conditions and performed field testing in calm waters using human-paddled kayaks as proxies for both divers and AUVs. Miller’s team then integrated these algorithms into a mission-relevant AUV and began testing them under more realistic ocean conditions, initially with a support boat acting as a diver surrogate, and then with actual divers.

„We quickly learned that you need more sensing capabilities on the diver when you factor in ocean currents,“ Miller explains. „With the algorithms demonstrated by MIT, the vehicle only needed to calculate the distance, or range, to the diver at regular intervals to solve the optimization problem of estimating the positions of both the vehicle and diver over time. But with the real ocean forces pushing everything around, this optimization problem blows up quickly.“

On the perception side, Miller’s team has been developing an AI classifier that can process both optical and sonar data mid-mission and solicit human input for any objects classified with uncertainty.

„The idea is for the classifier to pass along some information — say, a bounding box around an image — to the diver and indicate, „I think this is a tire, but I’m not sure. What do you think?“ Then, the diver can respond, „Yes, you’ve got it right, or no, look over here in the image to improve your classification,“ Miller says.

This feedback loop requires an underwater acoustic modem to support diver-AUV communication. State-of-the-art data rates in underwater acoustic communications would require tens of minutes to send an uncompressed image from the AUV to the diver. So, one aspect the team is investigating is how to compress information into a minimum amount to be useful, working within the constraints of the low bandwidth and high latency of underwater communications and the low size, weight, and power of the commercial off-the-shelf (COTS) hardware they’re using. For their prototype system, the team procured mostly COTS sensors and built a sensor payload that would easily integrate into an AUV routinely employed by the U.S. Navy, with the goal of facilitating technology transition. Beyond sonar and optical sensors, the payload features an acoustic modem for ranging to the diver and several data processing and compute boards.

Miller’s team has tested the sensor-equipped AUV and algorithms around coastal New England — including in the open ocean near Portsmouth, New Hampshire, with the University of New Hampshire’s (UNH) Gulf Surveyor and Gulf Challenger coastal research vessels as diver surrogates, and on the Boston-area Charles River, with an MIT Sailing Pavilion skiff as the surrogate.

„The UNH boats are well-equipped and can access realistic ocean conditions. But pretending to be a diver with a large boat is hard. With the skiff, we can move more slowly and get the relative motion in tune with how a diver and AUV would navigate together.“

Last summer, the team started testing equipment with human divers at Michigan Technological University’s Great Lakes Research Center. Although the divers lacked an interface to feed back information to the AUV, each swam holding the team’s tube-shaped prototype tablet, dubbed a „tube-let.“ The tube-let was equipped with a pressure and depth sensor, inertial measurement unit (to track relative motion), and ranging modem — all necessary components for the navigation algorithms to solve the optimization problem.

„A challenge during testing was coordinating the motion of the diver and vehicle, because they don’t yet collaborate,“ Miller says. „Once the divers go underwater, there is no communication with the team on the surface. So, you have to plan where to put the diver and vehicle so they don’t collide.“

The team also worked on the perception problem. The water clarity of the Great Lakes at that time of year allowed for underwater imaging with an optical sensor. Caroline Keenan, a Lincoln Scholars Program PhD student jointly working in the laboratory’s Advanced Undersea Systems and Technology Group and Leonard’s research group at MIT, took the opportunity to advance her work on knowledge transfer from optical sensors to sonar sensors. She is exploring whether optical classifiers can train sonar classifiers to recognize objects for which sonar data doesn’t exist. The motivation is to reduce the human operator load associated with labeling sonar data and training sonar classifiers.

With the internally funded research program coming to an end, Miller’s team is now seeking external sponsorship to refine and transition the technology to military or commercial partners.

„The modern world runs on undersea telecommunication and power cables, which are vulnerable to attack by disruptive actors. The undersea domain is becoming increasingly contested as more nations develop and advance the capabilities of autonomous maritime systems. Maintaining global economic security and U.S. strategic advantage in the undersea domain will require leveraging and combining the best of AI and human capabilities,“ Miller says.

Training a large artificial intelligence model is expensive, not just in dollars, but in time, energy, and computational resources. Traditionally, obtaining a smaller, faster model either requires training a massive one first and then trimming it down, or training a small one from scratch and accepting weaker performance. 

Researchers at MIT’s Computer Science and Artificial Intelligence Laboratory (CSAIL), Max Planck Institute for Intelligent Systems, European Laboratory for Learning and Intelligent Systems, ETH, and Liquid AI have now developed a new method that sidesteps this trade-off entirely, compressing models during training, rather than after.

The technique, called CompreSSM, targets a family of AI architectures known as state-space models, which power applications ranging from language processing to audio generation and robotics. By borrowing mathematical tools from control theory, the researchers can identify which parts of a model are pulling their weight and which are dead weight, before surgically removing the unnecessary components early in the training process.

„It’s essentially a technique to make models grow smaller and faster as they are training,“ says Makram Chahine, a PhD student in electrical engineering and computer science, CSAIL affiliate, and lead author of the paper. „During learning, they’re also getting rid of parts that are not useful to their development.“

The key insight is that the relative importance of different components within these models stabilizes surprisingly early during training. Using a mathematical quantity called Hankel singular values, which measure how much each internal state contributes to the model’s overall behavior, the team showed they can reliably rank which dimensions matter and which don’t after only about 10 percent of the training process. Once those rankings are established, the less-important components can be safely discarded, and the remaining 90 percent of training proceeds at the speed of a much smaller model.

„What’s exciting about this work is that it turns compression from an afterthought into part of the learning process itself,” says senior author Daniela Rus, MIT professor and director of CSAIL. “Instead of training a large model and then figuring out how to make it smaller, CompreSSM lets the model discover its own efficient structure as it learns. That’s a fundamentally different way to think about building AI systems.”

The results are striking. On image classification benchmarks, compressed models maintained nearly the same accuracy as their full-sized counterparts while training up to 1.5 times faster. A compressed model reduced to roughly a quarter of its original state dimension achieved 85.7 percent accuracy on the CIFAR-10 benchmark, compared to just 81.8 percent for a model trained at that smaller size from scratch. On Mamba, one of the most widely used state-space architectures, the method achieved approximately 4x training speedups, compressing a 128-dimensional model down to around 12 dimensions while maintaining competitive performance.

„You get the performance of the larger model, because you capture most of the complex dynamics during the warm-up phase, then only keep the most-useful states,“ Chahine says. „The model is still able to perform at a higher level than training a small model from the start.“

What makes CompreSSM distinct from existing approaches is its theoretical grounding. Conventional pruning methods train a full model and then strip away parameters after the fact, meaning you still pay the full computational cost of training the big model. Knowledge distillation, another popular technique, requires training a large „teacher“ model to completion and then training a second, smaller „student“ model on top of it, essentially doubling the training effort. CompreSSM avoids both of these costs by making informed compression decisions mid-stream.

The team benchmarked CompreSSM head-to-head against both alternatives. Compared to Hankel nuclear norm regularization, a recently proposed spectral technique for encouraging compact state-space models, CompreSSM was more than 40 times faster, while also achieving higher accuracy. The regularization approach slowed training by roughly 16 times because it required expensive eigenvalue computations at every single gradient step, and even then, the resulting models underperformed. Against knowledge distillation on CIFAR-10, CompressSM held a clear advantage for heavily compressed models: At smaller state dimensions, distilled models saw significant accuracy drops, while CompreSSM-compressed models maintained near-full performance. And because distillation requires a forward pass through both the teacher and student at every training step, even its smaller student models trained slower than the full-sized baseline.

The researchers proved mathematically that the importance of individual model states changes smoothly during training, thanks to an application of Weyl’s theorem, and showed empirically that the relative rankings of those states remain stable. Together, these findings give practitioners confidence that dimensions identified as negligible early on won’t suddenly become critical later.

The method also comes with a pragmatic safety net. If a compression step causes an unexpected performance drop, practitioners can revert to a previously saved checkpoint. „It gives people control over how much they’re willing to pay in terms of performance, rather than having to define a less-intuitive energy threshold,“ Chahine explains.

There are some practical boundaries to the technique. CompreSSM works best on models that exhibit a strong correlation between the internal state dimension and overall performance, a property that varies across tasks and architectures. The method is particularly effective on multi-input, multi-output (MIMO) models, where the relationship between state size and expressivity is strongest. For per-channel, single-input, single-output architectures, the gains are more modest, since those models are less sensitive to state dimension changes in the first place.

The theory applies most cleanly to linear time-invariant systems, although the team has developed extensions for the increasingly popular input-dependent, time-varying architectures. And because the family of state-space models extends to architectures like linear attention, a growing area of interest as an alternative to traditional transformers, the potential scope of application is broad.

Chahine and his collaborators see the work as a stepping stone. The team has already demonstrated an extension to linear time-varying systems like Mamba, and future directions include pushing CompreSSM further into matrix-valued dynamical systems used in linear attention mechanisms, which would bring the technique closer to the transformer architectures that underpin most of today’s largest AI systems.

„This had to be the first step, because this is where the theory is neat and the approach can stay principled,“ Chahine says. „It’s the stepping stone to then extend to other architectures that people are using in industry today.“

„The work of Chahine and his colleagues provides an intriguing, theoretically grounded perspective on compression for modern state-space models (SSMs),“ says Antonio Orvieto, ELLIS Institute Tübingen principal investigator and MPI for Intelligent Systems independent group leader, who wasn’t involved in the research. „The method provides evidence that the state dimension of these models can be effectively reduced during training and that a control-theoretic perspective can successfully guide this procedure. The work opens new avenues for future research, and the proposed algorithm has the potential to become a standard approach when pre-training large SSM-based models.“

The work, which was accepted as a conference paper at the International Conference on Learning Representations 2026, will be presented later this month. It was supported, in part, by the Max Planck ETH Center for Learning Systems, the Hector Foundation, Boeing, and the U.S. Office of Naval Research.

To improve data center efficiency, multiple storage devices are often pooled together over a network so many applications can share them. But even with pooling, significant device capacity remains underutilized due to performance variability across the devices.

MIT researchers have now developed a system that boosts the performance of storage devices by handling three major sources of variability simultaneously. Their approach delivers significant speed improvements over traditional methods that tackle only one source of variability at a time.

The system uses a two-tier architecture, with a central controller that makes big-picture decisions about which tasks each storage device performs, and local controllers for each machine that rapidly reroute data if that device is struggling.

The method, which can adapt in real-time to shifting workloads, does not require specialized hardware. When the researchers tested this system on realistic tasks like AI model training and image compression, it nearly doubled the performance delivered by traditional approaches. By intelligently balancing the workloads of multiple storage devices, the system can increase overall data center efficiency.

“There is a tendency to want to throw more resources at a problem to solve it, but that is not sustainable in many ways. We want to be able to maximize the longevity of these very expensive and carbon-intensive resources,” says Gohar Chaudhry, an electrical engineering and computer science (EECS) graduate student and lead author of a paper on this technique. “With our adaptive software solution, you can still squeeze a lot of performance out of your existing devices before you need to throw them away and buy new ones.”

Chaudhry is joined on the paper by Ankit Bhardwaj, an assistant professor at Tufts University; Zhenyuan Ruan PhD ’24; and senior author Adam Belay, an associate professor of EECS and a member of the MIT Computer Science and Artificial Intelligence Laboratory. The research will be presented at the USENIX Symposium on Networked Systems Design and Implementation.

Leveraging untapped performance

Solid-state drives (SSDs) are high-performance digital storage devices that allow applications to read and write data. For instance, an SSD can store vast datasets and rapidly send data to a processor for machine-learning model training.   

Pooling multiple SSDs together so many applications can share them improves efficiency, since not every application needs to use the entire capacity of an SSD at a given time. But not all SSDs perform equally, and the slowest device can limit the overall performance of the pool.

These inefficiencies arise from variability in SSD hardware and the tasks they perform.

To utilize this untapped SSD performance, the researchers developed Sandook, a software-based system that tackles three major forms of performance-hampering variability simultaneously. “Sandook” is an Urdu word that means “box,” to signify “storage.”

One type of variability is caused by differences in the age, amount of wear, and capacity of SSDs that may have been purchased at different times from multiple vendors.

The second type of variability is due to the mismatch between read and write operations occurring on the same SSD. To write new data to the device, the SSD must erase some existing data. This process can slow down data reads, or retrievals, happening at the same time.

The third source of variability is garbage collection, a process of gathering and removing outdated data to free up space. This process, which slows SSD operations, is triggered at random intervals that a data center operator cannot control.

“I can’t assume all SSDs will behave identically through my entire deployment cycle. Even if I give them all the same workload, some of them will be stragglers, which hurts the net throughput I can achieve,” Chaudhry explains.

Plan globally, react locally

To handle all three sources of variability, Sandook utilizes a two-tier structure. A global schedular optimizes the distribution of tasks for the overall pool, while faster schedulers on each SSD react to urgent events and shift operations away from congested devices.

The system overcomes delays from read-write interference by rotating which SSDs an application can use for reads and writes. This reduces the chance reads and writes happen simultaneously on the same machine.

Sandook also profiles the typical performance of each SSD. It uses this information to detect when garbage collection is likely slowing operations down. Once detected, Sandook reduces the workload on that SSD by diverting some tasks until garbage collection is finished.

“If that SSD is doing garbage collection and can’t handle the same workload anymore, I want to give it a smaller workload and slowly ramp things back up. We want to find the sweet spot where it is still doing some work, and tap into that performance,” Chaudhry says.

The SSD profiles also allow Sandook’s global controller to assign workloads in a weighted fashion that considers the characteristics and capacity of each device.

Because the global controller sees the overall picture and the local controllers react on the fly, Sandook can simultaneously manage forms of variability that happen over different time scales. For instance, delays from garbage collection occur suddenly, while latency caused by wear and tear builds up over many months.

The researchers tested Sandook on a pool of 10 SSDs and evaluated the system on four tasks: running a database, training a machine-learning model, compressing images, and storing user data. Sandook boosted the throughput of each application between 12 and 94 percent when compared to static methods, and improved the overall utilization of SSD capacity by 23 percent.

The system enabled SSDs to achieve 95 percent of their theoretical maximum performance, without the need for specialized hardware or application-specific updates.

“Our dynamic solution can unlock more performance for all the SSDs and really push them to the limit. Every bit of capacity you can save really counts at this scale,” Chaudhry says.

In the future, the researchers want to incorporate new protocols available on the latest SSDs that give operators more control over data placement. They also want to leverage the predictability in AI workloads to increase the efficiency of SSD operations.

“Flash storage is a powerful technology that underpins modern datacenter applications, but sharing this resource across workloads with widely varying performance demands remains an outstanding challenge. This work moves the needle meaningfully forward with an elegant and practical solution ready for deployment, bringing flash storage closer to its full potential in production clouds,” says Josh Fried, a software engineer at Google and incoming assistant professor at the University of Pennsylvania, who was not involved with this work.

This research was funded, in part, by the National Science Foundation, the U.S. Defense Advanced Research Projects Agency, and the Semiconductor Research Corporation.

Today, there are 94 nuclear reactors operating in the United States, more than in any other country in the world, and these units collectively provide nearly 20 percent of the nation’s electricity. That is a major accomplishment, according to Dean Price, but he believes that our country needs much more out of nuclear energy, especially at a moment when alternatives to fossil fuel-based power plants are desperately being sought. He became a nuclear engineer for this very reason — to make sure that nuclear technology is up to the task of delivering in this time of considerable need.

“Nuclear energy has been a tremendous part of our nation’s energy infrastructure for the past 60 years, and the number of people who maintain that infrastructure is incredibly small,” says Price, an MIT assistant professor in the Department of Nuclear Science and Engineering (NSE), as well as the Atlantic Richfield Career Development Professor in Energy Studies. “By becoming a nuclear engineer, you become one of a select number of people responsible for carbon-free energy generation in the United States.” 

That was a mission he was eager to take part in, and the goals he set for himself were far from modest: He wanted to help design and usher in a new class of nuclear reactors, building on the safety, economics, and reliability of the existing nuclear fleet.

Price has never wavered from this objective, and he’s only found encouragement along the way. The nuclear engineering community, he says, “is small, close-knit, and very welcoming. Once you get into it, most people are not inclined to do anything else.”

Illuminating the relationships between physical processes

In his first research project as an undergraduate at the University of Illinois Urbana at Champaign, Price studied the safety of the steel and concrete casks used to store spent reactor fuel rods after they’ve cooled off in tanks of water, typically for several years. His analysis indicated that this storage method was quite safe, although the question as to what should ultimately be done with these fuel casks, in terms of long-term disposal, remains open in this country.

After starting graduate studies at the University of Michigan in 2020, Price took up a different line of research that he’s still engaged in today. That area of study, called multiphysics modeling, involves looking at various physical processes going on in the core of a nuclear reactor to see how they interact — an alternative to studying these processes one at a time.

One key process, neutronics, concerns how neutrons buzz around in the reactor core causing nuclear fission, which is what generates the power. A second process, called thermal hydraulics, involves cooling the reactor to extract the heat generated by neutrons. A multiphysics simulation, analyzing how these two processes interact, could show how the heat carried away as the reactor produces power affects the behavior of neutrons, because the hotter the fuel is, the less likely it is to cause fission.

“If you ever want to change your power level, or do anything with the reactor, the temperature of the fuel is a critical input that you need to know,” says Price. “Multiphysics modeling allows us to correlate the fission neutronics processes with a thermal property, temperature. That, in turn, can help us predict how the reactor will behave under different conditions.”

Multiphysics modeling for light water reactors, which are the ones operating today with capacities on the order of 1,000 megawatts, are pretty well established, Prices says. But methods for modeling advanced reactors — small modular reactors (SMRs with capacities ranging from around 20 to 300 MW) and microreactors (rated at 1 to 20 MW) — are far less advanced. Only a very small number of these reactors are operating today, but Price is focusing his efforts on them because of their potential to produce power more cheaply and more safely, along with their greater flexibility in power and size.   

Although multiphysics simulations have supplied the nuclear community with a wealth of information, they can require supercomputers to solve, or find approximate solutions to, coupled and extremely difficult nonlinear equations. In the hopes of greatly reducing the computational burden, Price is actively exploring artificial intelligence approaches that could provide similar answers while bypassing those burdensome equations altogether. That has been a central theme of his research agenda since he joined the MIT faculty in September 2025.

A crucial role for artificial intelligence

What artificial intelligence and machine-learning methods, in particular, are good at is finding patterns concealed within data, such as correlations between variables critical to the functioning of a nuclear plant. For example, Price says, “if you tell me the power level of your reactor, it [AI] could tell you what the fuel temperature is and even tell you the 3-dimensional temperature distribution in your core.” And if this can be done without solving any complicated differential equations, computational costs could be greatly reduced.

Price is investigating several applications where AI may be especially useful, such as helping with the design of novel kinds of reactors. “We could then rely on the safety frameworks developed over the past 50 years to carry out a safety analysis of the proposed design,” he says. “In this way, AI will not be directly interfacing with anything that is safety-critical.” As he sees it, AI’s role would be to augment established procedures, rather than replacing them, helping to fill in existing gaps in knowledge.

When a machine-learning model is given a sufficient amount of data to learn from, it can help us better understand the relationship between key physical processes — again without having to solve nonlinear differential equations. 

“By really pinning down those relationships, we can make better design decisions in the early stages,” Price says. “And when that technology is developed and deployed, AI can help us make more intelligent control decisions that will enable us to operate our reactors in a safer and more economical way.”

Giving back to the community that nurtured him

Simply put, one of his chief goals is to bring the benefits of AI to the nuclear industry, and he views the possibilities as vast and largely untapped. Price also believes that he is well-positioned as a professor at MIT to bring us closer to the nuclear future that he envisions. As he sees it, he’s working not only to develop the next generation of reactors, but also to help prepare the next generation of leaders in the field.

Price became acquainted with some prospective members of that “next generation” in a design course he co-taught last fall with Curtis Smith, the KEPCO Professor of the Practice of Nuclear Science and Engineering. For Price, that introduction lasted just a few months, but it was long enough for him to discover that MIT students are exceptionally motivated, hard-working, and capable. Not surprisingly, those happen to be the same qualities he’s hoping to find in the students that join his research team.

Price vividly recalls the support he received when taking his first, tentative steps in this field. Now that he’s moved up the ranks from undergraduate to professor, and acquired a substantial body of knowledge along the way, he wants his students “to experience that same feeling that I had upon entering the field.” Beyond his specific goals for improving the design and operation of nuclear reactors, Price says, “I hope to perpetuate the same fun and healthy environment that made me love nuclear engineering in the first place.”

Artificial intelligence is increasingly being used to help optimize decision-making in high-stakes settings. For instance, an autonomous system can identify a power distribution strategy that minimizes costs while keeping voltages stable.

But while these AI-driven outputs may be technically optimal, are they fair? What if a low-cost power distribution strategy leaves disadvantaged neighborhoods more vulnerable to outages than higher-income areas?

To help stakeholders quickly pinpoint potential ethical dilemmas before deployment, MIT researchers developed an automated evaluation method that balances the interplay between measurable outcomes, like cost or reliability, and qualitative or subjective values, such as fairness.   

The system separates objective evaluations from user-defined human values, using a large language model (LLM) as a proxy for humans to capture and incorporate stakeholder preferences. 

The adaptive framework selects the best scenarios for further evaluation, streamlining a process that typically requires costly and time-consuming manual effort. These test cases can show situations where autonomous systems align well with human values, as well as scenarios that unexpectedly fall short of ethical criteria.

“We can insert a lot of rules and guardrails into AI systems, but those safeguards can only prevent the things we can imagine happening. It is not enough to say, ‘Let’s just use AI because it has been trained on this information.’ We wanted to develop a more systematic way to discover the unknown unknowns and have a way to predict them before anything bad happens,” says senior author Chuchu Fan, an associate professor in the MIT Department of Aeronautics and Astronautics (AeroAstro) and a principal investigator in the MIT Laboratory for Information and Decision Systems (LIDS).

Fan is joined on the paper by lead author Anjali Parashar, a mechanical engineering graduate student; Yingke Li, an AeroAstro postdoc; and others at MIT and Saab. The research will be presented at the International Conference on Learning Representations.

Evaluating ethics

In a large system like a power grid, evaluating the ethical alignment of an AI model’s recommendations in a way that considers all objectives is especially difficult.

Most testing frameworks rely on pre-collected data, but labeled data on subjective ethical criteria are often hard to come by. In addition, because ethical values and AI systems are both constantly evolving, static evaluation methods based on written codes or regulatory documents require frequent updates.

Fan and her team approached this problem from a different perspective. Drawing on their prior work evaluating robotic systems, they developed an experimental design framework to identify the most informative scenarios, which human stakeholders would then evaluate more closely.

Their two-part system, called Scalable Experimental Design for System-level Ethical Testing (SEED-SET), incorporates quantitative metrics and ethical criteria. It can identify scenarios that effectively meet measurable requirements and align well with human values, and vice versa.   

“We don’t want to spend all our resources on random evaluations. So, it is very important to guide the framework toward the test cases we care the most about,” Li says.

Importantly, SEED-SET does not need pre-existing evaluation data, and it adapts to multiple objectives.

For instance, a power grid may have several user groups, including a large rural community and a data center. While both groups may want low-cost and reliable power, each group’s priority from an ethical perspective may vary widely.

These ethical criteria may not be well-specified, so they can’t be measured analytically.

The power grid operator wants to find the most cost-effective strategy that best meets the subjective ethical preferences of all stakeholders.

SEED-SET tackles this challenge by splitting the problem into two, following a hierarchical structure. An objective model considers how the system performs on tangible metrics like cost. Then a subjective model that considers stakeholder judgements, like perceived fairness, builds on the objective evaluation.

“The objective part of our approach is tied to the AI system, while the subjective part is tied to the users who are evaluating it. By decomposing the preferences in a hierarchical fashion, we can generate the desired scenarios with fewer evaluations,” Parashar says.

Encoding subjectivity

To perform the subjective assessment, the system uses an LLM as a proxy for human evaluators. The researchers encode the preferences of each user group into a natural language prompt for the model.

The LLM uses these instructions to compare two scenarios, selecting the preferred design based on the ethical criteria.

“After seeing hundreds or thousands of scenarios, a human evaluator can suffer from fatigue and become inconsistent in their evaluations, so we use an LLM-based strategy instead,” Parashar explains.

SEED-SET uses the selected scenario to simulate the overall system (in this case, a power distribution strategy). These simulation results guide its search for the next best candidate scenario to test.

In the end, SEED-SET intelligently selects the most representative scenarios that either meet or are not aligned with objective metrics and ethical criteria. In this way, users can analyze the performance of the AI system and adjust its strategy.

For instance, SEED-SET can pinpoint cases of power distribution that prioritize higher-income areas during periods of peak demand, leaving underprivileged neighborhoods more prone to outages.

To test SEED-SET, the researchers evaluated realistic autonomous systems, like an AI-driven power grid and an urban traffic routing system. They measured how well the generated scenarios aligned with ethical criteria.

The system generated more than twice as many optimal test cases as the baseline strategies in the same amount of time, while uncovering many scenarios other approaches overlooked.

“As we shifted the user preferences, the set of scenarios SEED-SET generated changed drastically. This tells us the evaluation strategy responds well to the preferences of the user,” Parashar says.

To measure how useful SEED-SET would be in practice, the researchers will need to conduct a user study to see if the scenarios it generates help with real decision-making.

In addition to running such a study, the researchers plan to explore the use of more efficient models that can scale up to larger problems with more criteria, such as evaluating LLM decision-making.

This research was funded, in part, by the U.S. Defense Advanced Research Projects Agency.

Designers, makers, and others often use 3D printing to rapidly prototype a range of functional objects, from movie props to medical devices. Accurate print previews are essential so users know a fabricated object will perform as expected.

But previews generated by most 3D-printing software focus on function rather than aesthetics. A printed object may end up with a different color, texture, or shading than the user expected, resulting in multiple reprints that waste time, effort, and material.

To help users envision how a fabricated object will look, researchers from MIT and elsewhere developed an easy-to-use preview tool that puts appearance first.

Users upload a screenshot of the object from their 3D-printing software, along with a single image of the print material. From these inputs, the system automatically generates a rendering of how the fabricated object is likely to look.

The artificial intelligence-powered system, called VisiPrint, is designed to work with a range of 3D-printing software and can handle any material example. It considers not only the color of the material, but also gloss, translucency, and how nuances of the fabrication process affect the object’s appearance.

Such aesthetics-focused previews could be especially useful in areas like dentistry, by helping clinicians ensure temporary crowns and bridges match the appearance of a patient’s teeth, or in architecture, to aid designers in assessing the visual impact of models.

“3D printing can be a very wasteful process. Some studies estimate that as much as a third of the material used goes straight to the landfill, often from prototypes the user ends of discarding. To make 3D printing more sustainable, we want to reduce the number of tries it takes to get the prototype you want. The user shouldn’t have to try out every printing material they have before they settle on a design,” says Maxine Perroni-Scharf, an electrical engineering and computer science (EECS) graduate student and lead author of a paper on VisiPrint.

She is joined on the paper by Faraz Faruqi, a fellow EECS graduate student; Raul Hernandez, an MIT undergraduate; SooYeon Ahn, a graduate student at the Gwangju Institute of Science and Technology; Szymon Rusinkiewicz, a professor of computer science at Princeton University; William Freeman, the Thomas and Gerd Perkins Professor of EECS at MIT and a member of the Computer Science and Artificial Intelligence Laboratory (CSAIL); and senior author Stefanie Mueller, an associate professor of EECS and Mechanical Engineering at MIT, and a member of CSAIL. The research will be presented at the ACM CHI Conference on Human Factors in Computing Systems.

Accurate aesthetics

The researchers focused on fused deposition modeling (FDM), the most common type of 3D printing. In FDM, print material filament is melted and then squirted through a nozzle to fabricate an object one layer at a time.

Generating accurate aesthetic previews is challenging because the melting and extrusion process can change the appearance of a material, as can the height of each deposited layer and the path the nozzle follows during fabrication.

VisiPrint uses two AI models that work together to overcome those challenges.

The VisiPrint preview is based on two inputs: a screenshot of the digital design from a user’s 3D-printing software (called “slicer” software), and an image of the print material, which can be taken from an online source or captured from a printed sample.

From these inputs, a computer vision model extracts features from the material sample that are important for the object’s appearance.

It feeds those features to a generative AI model that computes the geometry and structure of the object, while incorporating the so-called “slicing” pattern the nozzle will follow as it extrudes each layer.

The key to the researchers’ approach is a special conditioning method. This involves carefully adjusting the inner workings of the model to guide it, so it follows the slicing pattern and obeys the constraints of the 3D-printing process.

Their conditioning method utilizes a depth map that preserves the shape and shading of the object, along with a map of the edges that reflects the internal contours and structural boundaries.

“If you don’t have the right balance of these two things, you could use up with bad geometry or an incorrect slicing pattern. We had to be careful to combine them in the right way,” Perroni-Scharf says.

A user-focused system

The team also produced an easy-to-use interface where one can upload the required images and evaluate the preview.

The VisiPrint interface enables more advanced makers to adjust multiple settings, such as the influence of certain colors on the final appearance.

In the end, the aesthetic preview is intended to complement the functional preview generated by slicer software, since VisiPrint does not estimate printability, mechanical feasibility, or likelihood of failure.

To evaluate VisiPrint, the researchers conducted a user study that asked participants to compare the system to other approaches. Nearly all participants said it provided better overall appearance as well as more textural similarity with printed objects.

In addition, the VisiPrint preview process took about a minute on average, which was more than twice as fast as any competing method.

“VisiPrint really shined when compared to other AI interfaces. If you give a more general AI model the same screenshots, it might randomly change the shape or use the wrong slicing pattern because it had no direct conditioning,” she says.

In the future, the researchers want to address artifacts that can occur when model previews have extremely fine details. They also want to add features that allow users to optimize parts of the printing process beyond color of the material.

“It is important to think about the way that we fabricate objects. We need to continue striving to develop methods that reduce waste. To that end, this marriage of AI with the physical making process is an exciting area of future work,” Perroni-Scharf says.

“‘What you see is what you get’ has been the main thing that made desktop publishing ‘happen’ in the 1980s, as it allowed users to get what they wanted at first try. It is time to get WYSIWYG for 3D printing as well. VisiPrint is a great step in this direction,” says Patrick Baudisch, a professor of computer science at the Hasso Plattner Institute, who was not involved with this work.

This research was funded, in part, by an MIT Morningside Academy for Design Fellowship and an MIT MathWorks Fellowship.

Billions of years ago, simple organic molecules drifted across Earth’s primordial landscape — nothing more than basic chemical compounds. But as natural forces shaped the planet over hundreds of millions of years, these molecules began to interact and bond in increasingly complex ways. Along the way, something spectacular emerged: life.

“Life is, to some degree, magical,” says computational biologist Sergei Kotelnikov. Simple organic compounds congregate into polymers, which assemble into living cells and ultimately organisms — the whole being greater than the sum of its parts.

“You can write formulas on how a molecule behaves,” he says, referring to the world of quantum mechanics. “But yet somehow, a few orders of magnitude above, on a bigger scale, it gives rise to such a mystery.”

Kotelnikov builds models to analyze and predict the structure of these biomolecules, particularly proteins, the fundamental building blocks of every organism. This year, he joined MIT as part of the School of Science Dean’s Postdoctoral Fellowship to work with the Keating Lab, where researchers focus on protein structure, function, and interaction. Using machine learning, his goal is to develop new methods in protein modeling with potential applications that span from medicine to agriculture.

A hunger for problems to solve

Kotelnikov grew up in Abakan, Russia, a small city sitting right in the center of Eurasia. As a child, one of his favorite pastimes was playing with Lego bricks.

“It encouraged me to build new things, rather than just following instructions,” he says. “You can do anything.”

Kotelnikov’s father, whose background lies in engineering and economics, would often challenge him with math problems.

“Your brain — you can feel some kind of expansion of understanding how things work, and that’s a very satisfactory feeling,” Kotelnikov says.

This itch to solve problems led him to join science Olympiad competitions, and later, a science-focused public boarding school located near the Russian Academy of Sciences, from which he often encountered scientists.

“It was like a candy shop,” he recalls, describing the period as a life-changing experience.

In 2012, Kotelnikov began his bachelor of science in physics and applied mathematics at the Moscow Institute of Physics and Technology — considered one of the leading STEM universities in Russia, and globally — and continued there for his master’s degree. It was there that biology came into the picture.

During a course on statistical physics, Kotelnikov was first introduced to the idea of the “emergence of complexity.” He became fascinated by this “mysterious and attractive manifestation of biology … this evolution that sharpens the physical phenomenon” to create, drive, and shape life as we know it today. By the time he completed his master’s degree, he realized he had only scratched surface of the field of computational biology.

In 2018, he began his PhD at Stony Brook University in New York and began working with Dima Kozakov, who is recognized as one of the world’s leaders in predicting protein interactions and complex structures.

Studying the architecture of life  

Proteins act like the bricks that construct an organism, underpinning almost every cellular process from tissue repair to hormone production. Like pieces of a Lego tower, their structures and interactions determine the functions that they carry out in a body.

However, diseases arise when they’re folded, curled, twisted, or connected in unusual ways. To develop medical interventions, scientists break down the tower and examine each individual piece to find the culprit and correct their shape and pairing. With limited experimental data on protein structures and interactions currently available, simulations developed by computational biologists like Kotelnikov provide crucial insight that inform fundamental understanding and applications like drug discovery.

With the guidance of Kozakov at Stony Brook’s Laufer Center for Physical and Quantitative Biology, Kotelnikov carried over his understanding of physics to create modeling methods that are more effective, efficient, reliable, and generalizable. Among them, he developed a new way of predicting the protein complex structures mediated by proteolysis-targeting chimeras, or PROTACs, a new class of molecules that can trigger the breakdown of specific proteins previously considered undruggable, such as those found in cancer.

PROTACs have been challenging to model, in part because they are composed of proteins that don’t naturally interact with each other, and because the linker that connects them is flexible. Imagine trying to guess the overall shape of a bendy Lego piece attached to two other pieces of different irregular, unmatched shapes. To efficiently find all possible configurations, Kotelnikov’s method conceptually cuts the linker into two halves and models each separately, then reformulates the problem and calculates it using a powerful algorithm called Fast Fourier Transform.

“It’s kind of like applied math judo that you sometimes need to do in order to make certain intractable computations tractable,” he says.

Kotelnikov’s state-of-the-art methods have been instrumental to his team’s top performance in numerous international challenges including the Critical Assessment of protein Structure Prediction (CASP) competition — the same contest in which the Nobel Prize-winning AlphaFold system for protein 3D structure prediction was presented.

Physics and machine learning

At MIT, Kotelnikov is working with Amy Keating, the Jay A. Stein (1968) Professor of Biology, biology department head, and professor of biological engineering, to study protein structure, function, and interactions.  

A recognized leader in the field, Keating employs both computational and experimental methods to study proteins, their interactions, as well as how this can impact disease. By infusing physics with machine learning, Kotelnikov’s goal is to advance modeling methods that can vastly inform applications such as cancer immunology and crop protection.

“Kotelnikov stands to gain a lot from working closely with wet lab researchers who are doing the experiments that will complement and test his predictions, and my lab will benefit from his experience developing and applying advanced computational analyses,” says Keating.

Kotelnikov is also planning to work with professors Tommi Jaakkola and Tess Smidt in MIT’s Department of Electrical Engineering and Computer Science to explore a field called geometric deep learning. In particular, he aims to integrate physical and geometric knowledge about biomolecules into neural network architectures and learning procedures. This approach can significantly reduce the amount of data needed for learning, and improve the generalizability of resulting models.

Beyond the two departments, Kotelnikov is also excited to see how the diversity and interdisciplinary mix of MIT’s community will help him come up with ideas.

“When you’re building a model, you’re entering this imaginary world of assumptions and simplifications and it might feel challenging because of this disconnect with reality,” Kotelnikov says. “Being able to efficiently communicate with experimentalists is of high value.”

In biology, defects are generally bad. But in materials science, defects can be intentionally tuned to give materials useful new properties. Today, atomic-scale defects are carefully introduced during the manufacturing process of products like steel, semiconductors, and solar cells to help improve strength, control electrical conductivity, optimize performance, and more.

But even as defects have become a powerful tool, accurately measuring different types of defects and their concentrations in finished products has been challenging, especially without cutting open or damaging the final material. Without knowing what defects are in their materials, engineers risk making products that perform poorly or have unintended properties.

Now, MIT researchers have built an AI model capable of classifying and quantifying certain defects using data from a noninvasive neutron-scattering technique. The model, which was trained on 2,000 different semiconductor materials, can detect up to six kinds of point defects in a material simultaneously, something that would be impossible using conventional techniques alone.

“Existing techniques can’t accurately characterize defects in a universal and quantitative way without destroying the material,” says lead author Mouyang Cheng, a PhD candidate in the Department of Materials Science and Engineering. “For conventional techniques without machine learning, detecting six different defects is unthinkable. It’s something you can’t do any other way.”

The researchers say the model is a step toward harnessing defects more precisely in products like semiconductors, microelectronics, solar cells, and battery materials.

“Right now, detecting defects is like the saying about seeing an elephant: Each technique can only see part of it,” says senior author and associate professor of nuclear science and engineering Mingda Li. “Some see the nose, others the trunk or ears. But it is extremely hard to see the full elephant. We need better ways of getting the full picture of defects, because we have to understand them to make materials more useful.”

Joining Cheng and Li on the paper are postdoc Chu-Liang Fu, undergraduate researcher Bowen Yu, master’s student Eunbi Rha, PhD student Abhijatmedhi Chotrattanapituk ’21, and Oak Ridge National Laboratory staff members Douglas L Abernathy PhD ’93 and Yongqiang Cheng. The paper appears today in the journal Matter.

Detecting defects

Manufacturers have gotten good at tuning defects in their materials, but measuring precise quantities of defects in finished products is still largely a guessing game.

“Engineers have many ways to introduce defects, like through doping, but they still struggle with basic questions like what kind of defect they’ve created and in what concentration,” Fu says. “Sometimes they also have unwanted defects, like oxidation. They don’t always know if they introduced some unwanted defects or impurity during synthesis. It’s a longstanding challenge.”

The result is that there are often multiple defects in each material. Unfortunately, each method for understanding defects has its limits. Techniques like X-ray diffraction and positron annihilation characterize only some types of defects. Raman spectroscopy can discern the type of defect but can’t directly infer the concentration. Another technique known as transmission electron microscope requires people to cut thin slices of samples for scanning.

In a few previous papers, Li and collaborators applied machine learning to experimental spectroscopy data to characterize crystalline materials. For the new paper, they wanted to apply that technique to defects.

For their experiment, the researchers built a computational database of 2,000 semiconductor materials. They made sample pairs of each material, with one doped for defects and one left without defects, then used a neutron-scattering technique that measures the different vibrational frequencies of atoms in solid materials. They trained a machine-learning model on the results.

“That built a foundational model that covers 56 elements in the periodic table,” Cheng says. “The model leverages the multihead attention mechanism, just like what ChatGPT is using. It similarly extracts the difference in the data between materials with and without defects and outputs a prediction of what dopants were used and in what concentrations.”

The researchers fine-tuned their model, verified it on experimental data, and showed it could measure defect concentrations in an alloy commonly used in electronics and in a separate superconductor material.

The researchers also doped the materials multiple times to introduce multiple point defects and test the limits of the model, ultimately finding it can make predictions about up to six defects in materials simultaneously, with defect concentrations as low as 0.2 percent.

“We were really surprised it worked that well,” Cheng says. “It’s very challenging to decode the mixed signals from two different types of defects — let alone six.”

A model approach

Typically, manufacturers of things like semiconductors run invasive tests on a small percentage of products as they come off the manufacturing line, a slow process that limits their ability to detect every defect.

“Right now, people largely estimate the quantities of defects in their materials,” Yu says. “It is a painstaking experience to check the estimates by using each individual technique, which only offers local information in a single grain anyway. It creates misunderstandings about what defects people think they have in their material.”

The results were exciting for the researchers, but they note their technique measuring the vibrational frequencies with neutrons would be difficult for companies to quickly deploy in their own quality-control processes.

“This method is very powerful, but its availability is limited,” Rha says. “Vibrational spectra is a simple idea, but in certain setups it’s very complicated. There are some simpler experimental setups based on other approaches, like Raman spectroscopy, that could be more quickly adopted.”

Li says companies have already expressed interest in the approach and asked when it will work with Raman spectroscopy, a widely used technique that measures the scattering of light. Li says the researchers’ next step is training a similar model based on Raman spectroscopy data. They also plan to expand their approach to detect features that are larger than point defects, like grains and dislocations.

For now, though, the researchers believe their study demonstrates the inherent advantage of AI techniques for interpreting defect data.

“To the human eye, these defect signals would look essentially the same,” Li says. “But the pattern recognition of AI is good enough to discern different signals and get to the ground truth. Defects are this double-edged sword. There are many good defects, but if there are too many, performance can degrade. This opens up a new paradigm in defect science.”

The work was supported, in part, by the Department of Energy and the National Science Foundation.