Aurelia Butler

Experimental computer memories and processors built from magnetic materials use far less energy than traditional silicon-based devices. Two-dimensional magnetic materials, composed of layers that are only a few atoms thick, have incredible properties that could allow magnetic-based devices to achieve unprecedented speed, efficiency, and scalability.
While many hurdles must be overcome until these so-called van der Waals magnetic materials can be integrated into functioning computers, MIT researchers took an important step in this direction by demonstrating precise control of a van der Waals magnet at room temperature.
This is key, since magnets composed of atomically thin van der Waals materials can typically only be controlled at extremely cold temperatures, making them difficult to deploy outside a laboratory.
The researchers used pulses of electrical current to switch the direction of the device’s magnetization at room temperature. Magnetic switching can be used in computation, the same way a transistor switches between open and closed to represent 0s and 1s in binary code, or in computer memory, where switching enables data storage.
The team fired bursts of electrons at a magnet made of a new material that can sustain its magnetism at higher temperatures. The experiment leveraged a fundamental property of electrons known as spin, which makes the electrons behave like tiny magnets. By manipulating the spin of electrons that strike the device, the researchers can switch its magnetization.
“The heterostructure device we have developed requires an order of magnitude lower electrical current to switch the van der Waals magnet, compared to that required for bulk magnetic devices,” says Deblina Sarkar, the AT&T Career Development Assistant Professor in the MIT Media Lab and Center for Neurobiological Engineering, head of the Nano-Cybernetic Biotrek Lab, and the senior author of a paper on this technique. “Our device is also more energy efficient than other van der Waals magnets that are unable to switch at room temperature.”
In the future, such a magnet could be used to build faster computers that consume less electricity. It could also enable magnetic computer memories that are nonvolatile, which means they don’t leak information when powered off, or processors that make complex AI algorithms more energy-efficient.

“There is a lot of inertia around trying to improve materials that worked well in the past. But we have shown that if you make radical changes, starting by rethinking the materials you are using, you can potentially get much better solutions,” says Shivam Kajale, a graduate student in Sarkar’s lab and co-lead author of the paper.
Kajale and Sarkar are joined on the paper by co-lead author Thanh Nguyen, a graduate student in the Department of Nuclear Science and Engineering (NSE); Corson Chao, a graduate student in the Department of Materials Science and Engineering (DSME); David Bono, a DSME research scientist; Artittaya Boonkird, an NSE graduate student; and Mingda Li, associate professor of nuclear science and engineering. The research appears this week in Nature Communications.
An atomically thin advantage
Methods to fabricate tiny computer chips in a clean room from bulk materials like silicon can hamper devices. For instance, the layers of material may be barely 1 nanometer thick, so minuscule rough spots on the surface can be severe enough to degrade performance.
By contrast, van der Waals magnetic materials are intrinsically layered and structured in such a way that the surface remains perfectly smooth, even as researchers peel off layers to make thinner devices. In addition, atoms in one layer won’t leak into other layers, enabling the materials to retain their unique properties when stacked in devices.
“In terms of scaling and making these magnetic devices competitive for commercial applications, van der Waals materials are the way to go,” Kajale says.

But there’s a catch. This new class of magnetic materials have typically only been operated at temperatures below 60 kelvins (-351 degrees Fahrenheit). To build a magnetic computer processor or memory, researchers need to use electrical current to operate the magnet at room temperature.
To achieve this, the team focused on an emerging material called iron gallium telluride. This atomically thin material has all the properties needed for effective room temperature magnetism and doesn’t contain rare earth elements, which are undesirable because extracting them is especially destructive to the environment.
Nguyen carefully grew bulk crystals of this 2D material using a special technique. Then, Kajale fabricated a two-layer magnetic device using nanoscale flakes of iron gallium telluride underneath a six-nanometer layer of platinum.
Tiny device in hand, they used an intrinsic property of electrons known as spin to switch its magnetization at room temperature.
Electron ping-pong
While electrons don’t technically “spin” like a top, they do possess the same kind of angular momentum. That spin has a direction, either up or down. The researchers can leverage a property known as spin-orbit coupling to control the spins of electrons they fire at the magnet.
The same way momentum is transferred when one ball hits another, electrons will transfer their “spin momentum” to the 2D magnetic material when they strike it. Depending on the direction of their spins, that momentum transfer can reverse the magnetization.
In a sense, this transfer rotates the magnetization from up to down (or vice-versa), so it is called a “torque,” as in spin-orbit torque switching. Applying a negative electric pulse causes the magnetization to go downward, while a positive pulse causes it to go upward.
The researchers can do this switching at room temperature for two reasons: the special properties of iron gallium telluride and the fact that their technique uses small amounts of electrical current. Pumping too much current into the device would cause it to overheat and demagnetize.
The team faced many challenges over the two years it took to achieve this milestone, Kajale says. Finding the right magnetic material was only half the battle. Since iron gallium telluride oxidizes quickly, fabrication must be done inside a glovebox filled with nitrogen.
“The device is only exposed to air for 10 or 15 seconds, but even after that I have to do a step where I polish it to remove any oxide,” he says.
Now that they have demonstrated room-temperature switching and greater energy efficiency, the researchers plan to keep pushing the performance of magnetic van der Waals materials.
“Our next milestone is to achieve switching without the need for any external magnetic fields. Our aim is to enhance our technology and scale up to bring the versatility of van der Waals magnet to commercial applications,” Sarkar says.
This work was carried out, in part, using the facilities at MIT.Nano and the Harvard University Center for Nanoscale Systems. More

Medical summarization, a process that uses artificial intelligence (AI) to condense complex patient information, is currently used in health care settings for tasks such as creating electronic health records and simplifying medical text for insurance claims processing. While the practice is intended to create efficiencies, it can be labor-intensive, according to Penn State researchers, who created a new method to streamline the way AI creates these summaries, efficiently producing more reliable results.
In their work, which was presented at the Proceedings of the 2023 Conference on Empirical Methods in Natural Language Processing in Singapore last December, the researchers introduced a framework to fine-tune the training of natural language processing (NLP) models that are used to create medical summaries.
“There is a faithfulness issue with the current NLP tools and machine learning algorithms used in medical summarization,” said Nan Zhang, a graduate student pursing a doctorate in informatics the College of Information Sciences and Technology (IST) and the first author on the paper. “To ensure records of doctor-patient interactions are reliable, a medical summarization model should remain 100% consistent with the reports and conversations they are documenting.”
Existing medical text summarization tools involve human supervision to prevent the generation of unreliable summaries that could lead to serious health care risks, according to Zhang. This “unfaithfulness” has been understudied despite its importance for ensuring safety and efficiency in healthcare reporting.
The researchers began by examining three datasets — online health question summarization, radiology report summarization and medical dialogue summarization — generated by existing AI models. They randomly selected between 100 and 200 summaries from each dataset and manually compared them to the doctors’ original medical reports, or source text, from which they were condensed. Summaries that did not accurately reflect the source text were placed into error categories.
“There are various types of errors that can occur with models that generate text,” Zhang said. “The model may miss a medical term or change it to something else. Summarization that is untrue or not consistent with source inputs can potentially cause harm to a patient.”
The data analysis revealed instances of summarization that were contradictory to the source text. For example, a doctor prescribed a medication to be taken three times a day, but the summary reported that the patient should not take said medication. The datasets also included what Zhang called “hallucinations,” resulting in summaries that contained extraneous information not supported by the source text.

The researchers set out to mitigate the unfaithfulness problem with their Faithfulness for Medical Summarization (FaMeSumm) framework. They began by using simple problem-solving techniques to construct sets of contrastive summaries — a set of faithful, error-free summaries and a set of unfaithful summaries containing errors. They also identified medical terms through external knowledge graphs or human annotations. Then, they fine-tuned existing pre-trained language models to the categorized data, modified objective functions to learn from the contrastive summaries and medical terms and made sure the models were trained to address each type of error instead of just mimicking specific words.
“Medical summarization models are trained to pay more attention to medical terms,” Zhang said. “But it’s important that those medical terms be summarized precisely as intended, which means including non-medical words like no, not or none. We don’t want the model to make modifications near or around those words, or the error is likely to be higher.”
FaMeSumm effectively and accurately summarized information from different kinds of training data. For example, if the provided training data comprised doctor notes, then the trained AI product was suited to generate summaries that facilitate doctors’ understanding of their notes. If the training data contained complex questions from patients, the trained AI product generated summaries that helped both patients and doctors understand the questions.
“Our method works on various kinds of datasets involving medical terms and for the mainstream, pre-trained language models we tested,” Zhang said. “It delivered a consistent improvement in faithfulness, which was confirmed by the medical doctors who checked our work.”
Fine-tuning large language models (LLMs) can be expensive and unnecessary, according to Zhang, so the experiments were conducted on five smaller mainstream language models.
“We did compare one of our fine-tuned models against GPT-3, which is an example of a large language model,” he said. “We found that our model reached significantly better performance in terms of faithfulness and showed the strong capability of our method, which is promising for its use on LLMs.”
This work contributes to the future of automated medical summarization, according to Zhang.

“Maybe, in the near future, AI will be trained to generate medical summaries as templates,” he said. “Doctors could simply doublecheck the output and make minor edits, which could significantly reduce the amount of time it takes to create the summaries.”
Prasenjit Mitra, professor in the College of IST and Zhang’s graduate adviser; Rui Zhang, assistant professor in the College of Engineering and Zhang’s graduate co-adviser; and Yusen Zhang, doctoral student in the College of Engineering — all from Penn State — and Wu Guo, with the Children’s Hospital Affiliated to Zhengzhou University in China, contributed to this research.
The Federal Ministry of Education and Research in Germany, under the LeibnizKILabor project, partially funded this research. Rui Zhang supported the travel funding. More

Think big. Despite its research topic, this could well be the motto of the Graphene Flagship, which was launched in 2013: With an overall budget of one billion Euros, it was Europe’s largest research initiative to date, alongside the Human Brain Flagship, which was launched at the same time. The same applies to the review article on the effects of graphene and related materials on health and the environment, which Empa researchers Peter Wick and Tina Bürki just published together with 30 international colleagues in the scientific journal ACS Nano; on 57 pages, they summarize the findings on the health and ecological risks of graphene materials, the reference list includes almost 500 original publications.
A wealth of knowledge — which also gives the all-clear. “We have investigated the potential acute effects of various graphene and graphene-like materials on the lungs, in the gastrointestinal tract and in the placenta — and no serious acute cell-damaging effects were observed in any of the studies,” says Wick, summarizing the results. Although stress reactions can certainly occur in lung cells, the tissue recovers rather quickly. However, some of the newer 2D materials such as boron nitrides, transition metal dichalcogenides, phosphenes and MXenes have not yet been investigated much, Wick points out; further investigations were needed here.
In their analyses, Wick and Co. did not limit themselves to newly produced graphene-like materials, but also looked at the entire life cycle of various applications of graphene-containing materials. In other words, they investigated questions such as: What happens when these materials are abraded or burnt? Are graphene particles released, and can this fine dust harm cells, tissues or the environment?
One example: The addition of a few percent graphene to polymers, such as epoxy resins or polyamides, significantly improves material properties such as mechanical stability or conductivity, but the abrasion particles do not cause any graphene-specific nanotoxic effect on the cells and tissues tested. Wick’s team will be able to continue this research even after the flagship project has come to an end, also thanks to funding from the EU as part of so-called Spearhead projects, of which Wick is deputy head.
In addition to Wick’s team, Empa researchers led by Bernd Nowack have used material flow analyses as part of the Graphene Flagship to calculate the potential future environmental impact of materials containing graphene and have modeled which ecosystems are likely to be impacted and to what extent. Roland Hischier’s team, like Nowack’s at Empa’s Technology and Society lab, used life cycle assessments to investigate the environmental sustainability of different production methods and application examples for various graphene-containing materials. And Roman Fasel’s team from Empa’s nanotech@surfaces lab has advanced the development of electronic components based on narrow graphene ribbons.
A European success story for research and innovation
Launched in 2013, the Graphene Flagship represented a completely new form of joint, coordinated research on an unprecedented scale. The aim of the large-scale project was to bring together researchers from research institutions and industry to bring practical applications based on graphene from the laboratory to the market within ten years, thereby creating economic growth, new jobs and new opportunities for Europe in key technologies. Over its ten-year lifetime, the consortium consisted of more than 150 academic and industrial research teams in 23 countries plus numerous associated members.

Last September, the ten-year funding period ended with the Graphene Week in Gothenburg, Sweden. The final report impressively demonstrates the success of the ambitious large-scale project: The Flagship has “produced” almost 5,000 scientific publications and more than 80 patents. It has created 17 spin-offs in the graphene sector, which have raised a total of more than 130 million Euros in venture capital. According to a study by the German economic research institute WifOR, the Graphene Flagship has led to a total added value of around 5.9 billion Euros in the participating countries and created more than 80,000 new jobs in Europe. This means that the impact of the Graphene Flagship is more than 10 times greater than shorter EU projects.
In the course of the project, Empa received a total of around three million Swiss francs in funding — which had a “catalytic” effect, as Peter Wick emphasizes: “We have roughly tripled this sum through follow-up projects totaling around 5.5 million Swiss francs, including further EU projects, projects funded by the Swiss National Science Foundation (SNSF) and direct cooperation projects with our industrial partners — and all this in the last five years.”
But the advantage of such projects goes far beyond the generous funding, emphasizes Wick: “It is truly unique to be involved in such a large project and broad network over such a long period of time. On the one hand, it has resulted in numerous new collaborations and ideas for projects. On the other hand, working together with international partners over such a long period of time has a completely different quality, we trust each other almost blindly; and such a well-coordinated team is much more efficient and produces better scientific results,” Wick is convinced. Last but not least, many personal friendships came about.
A new dimension: graphene and other 2D materials
Graphene is an enormously promising material. It consists of a single layer of carbon atoms arranged in a honeycomb pattern and has extraordinary properties: exceptional mechanical strength, flexibility, transparency and outstanding thermal and electrical conductivity. If the already two-dimensional material is spatially restricted even more, for example into a narrow ribbon, controllable quantum effects can be created. This could enable a wide range of applications, from vehicle construction and energy storage to quantum computing.
For a long time, this “miracle material” existed only in theory. It was not until 2004 that physicists Konstantin Novoselov and Andre Geim at the University of Manchester were able to specifically produce and characterize graphene. To do this, the researchers removed layers of graphite with a piece of adhesive tape until they had flakes just one atom thick. They were awarded the Nobel Prize in Physics for this work in 2010.
Since then, graphene has been the subject of intensive research. In the meantime, researchers have discovered more 2D materials, such as graphene-derived graphene acid, graphene oxide and cyanographs, which could have applications in medicine. Researchers want to use inorganic 2D materials such as boron nitride or MXenes to build batteries that are more powerful, develop electronic components or improve other materials. More

Imagine doubling the processing power of your smartphone, tablet, personal computer, or server using the existing hardware already in these devices.
Hung-Wei Tseng, a UC Riverside associate professor of electrical and computer engineering, has laid out a paradigm shift in computer architecture to do just that in a recent paper titled, “Simultaneous and Heterogeneous Multithreading.”
Tseng explained that today’s computer devices increasingly have graphics processing units (GPUs), hardware accelerators for artificial intelligence (AI) and machine learning (ML), or digital signal processing units as essential components. These components process information separately, moving information from one processing unit to the next, which in effect creates a bottleneck.
In their paper, Tseng and UCR computer science graduate student Kuan-Chieh Hsu introduce what they call “simultaneous and heterogeneous multithreading” or SHMT. They describe their development of a proposed SHMT framework on an embedded system platform that simultaneously uses a multi-core ARM processor, an NVIDIA GPU, and a Tensor Processing Unit hardware accelerator.
The system achieved a 1.96 times speedup and a 51% reduction in energy consumption.
“You don’t have to add new processors because you already have them,” Tseng said.
The implications are huge.
Simultaneous use of existing processing components could reduce computer hardware costs while also reducing carbon emissions from the energy produced to keep servers running in warehouse-size data processing centers. It also could reduce the need for scarce freshwater used to keep servers cool.
Tseng’s paper, however, cautions that further investigation is needed to answer several questions about system implementation, hardware support, code optimization, and what kind of applications stand to benefit the most, among other issues.
The paper was presented at the 56th Annual IEEE/ACM International Symposium on Microarchitecture held in October in Toronto, Canada. The paper garnered recognition from Tseng’s professional peers in the Institute of Electrical and Electronics Engineers, or IEEE, who selected it as one of 12 papers included in the group’s “Top Picks from the Computer Architecture Conferences” issue to be published this coming summer. More

Researchers from Carnegie Mellon University, the University Hospital Bonn and the University of Bonn have created an open-source platform known as A-SOiD that can learn and predict user-defined behaviors, just from video. The results of the study have now been published in the journal Nature Methods.
“This technique works great at learning classifications for a variety of animal and human behaviors,” said Eric Yttri, Eberly Family Associate Professor of Biological Sciences at Carnegie Mellon. “This would not only work on behavior but also the behavior of anything if there are identifiable patterns: stock markets, earthquakes, proteomics. It’s a powerful pattern recognition machine.”
Unlike many artificial intelligence (AI) programs, A-SOiD is not a black box. Instead, the researchers allowed the program to re-learn what it did wrong. They first trained the program with a fraction of the dataset, with a focus on the program’s weaker beliefs. If the program was not certain, the algorithm would reinforce the belief of that training data.
Because A-SOiD was taught to focus on the algorithm’s uncertainty rather than treating all data the same, Alex Hsu, a recent Ph.D. alumnus from Carnegie Mellon, said that it avoids common biases found in other AI models.
AI tool does justice to every class in a data set
“It’s a different way of feeding data in,” Hsu said. “Usually, people go in with the entire data set of whatever behaviors they’re looking for. They rarely understand that the data can be imbalanced, meaning there could be a well-represented behavior in their set and a poorly represented behavior in their set. This bias could then propagate from the prediction process to the experimental findings. Our algorithm takes care of data balancing by only learning from weaker. Our method is better at fairly representing every class in a data set.”
Because A-SOiD is trained in a supervised fashion, it can be very precise. If given a dataset, it can determine the difference between a person’s normal shiver and the tremors of a patient with Parkinson’s disease. It also serves as a complementary method to their unsupervised behavior segmentation platform, B-SOiD, released two years ago.

Besides being an effective program, A-SOiD is highly accessible, capable of running on a normal computer and is available as open source on GitHub.
A-SOiD is accessible for everyone in science
Jens Tillmann, a postdoctoral researcher from the University of Bonn at the University Hospital Bonn, said that the idea of having this program open to all researchers was part of its impact.
“This project wouldn’t have been possible without the open science mindset that both of our labs, but also the entire community of neuroethology have shown in recent years,” Tillmann said. “I am excited to be part of this community and look forward to future collaborative projects with other experts in the field.”
Yttri and Martin K. Schwarz, principal investigator at the University Hospital Bonn and member of the Transdisciplinary Research Areas (TRA) “Life & Health” at the University of Bonn, plan on using A-SOiD in their own labs to further investigate the relationship between the brain and behavior. Yttri plans to use A-SOiD in conjunction with other tools to investigate the neural mechanisms underlying spontaneous behaviors. Schwartz will use A-SOiD in conjunction with other behavioral modalities for a fine-grained analysis of known behaviors in social interactions.
Both Yttri and Schwarz said they hope that A-SOiD will be used by other researchers across disciplines and countries.
“A-SOiD is an important development allowing an AI-based entry into behavioral classification and thus an excellent unique opportunity to better understand the causal relationship between brain activity and behavior,” Schwarz said. “We also hope that the development of A-SOiD will serve as an efficient trigger for forthcoming collaborative research projects focusing on behavioral research in Europe but also across the Atlantic.” More

Pollsters trying to predict presidential election results and physicists searching for distant exoplanets have at least one thing in common: They often use a tried-and-true scientific technique called Bayesian inference.
Bayesian inference allows these scientists to effectively estimate some unknown parameter — like the winner of an election — from data such as poll results. But Bayesian inference can be slow, sometimes consuming weeks or even months of computation time or requiring a researcher to spend hours deriving tedious equations by hand.
Researchers from MIT and elsewhere have introduced an optimization technique that speeds things up without requiring a scientist to do a lot of additional work. Their method can achieve more accurate results faster than another popular approach for accelerating Bayesian inference.
Using this new automated technique, a scientist could simply input their model and then the optimization method does all the calculations under the hood to provide an approximation of some unknown parameter. The method also offers reliable uncertainty estimates that can help a researcher understand when to trust its predictions.
This versatile technique could be applied to a wide array of scientific quandaries that incorporate Bayesian inference. For instance, it could be used by economists studying the impact of microcredit loans in developing nations or sports analysts using a model to rank top tennis players.
“When you actually dig into what people are doing in the social sciences, physics, chemistry, or biology, they are often using a lot of the same tools under the hood. There are so many Bayesian analyses out there. If we can build a really great tool that makes these researchers lives easier, then we can really make a difference to a lot of people in many different research areas,” says senior author Tamara Broderick, an associate professor in MIT’s Department of Electrical Engineering and Computer Science (EECS) and a member of the Laboratory for Information and Decision Systems and the Institute for Data, Systems, and Society.
Broderick is joined on the paper by co-lead authors Ryan Giordano, an assistant professor of statistics at the University of California at Berkeley; and Martin Ingram, a data scientist at the AI company KONUX. The paper was recently published in the Journal of Machine Learning Research.
Faster results

When researchers seek a faster form of Bayesian inference, they often turn to a technique called automatic differentiation variational inference (ADVI), which is often both fast to run and easy to use.
But Broderick and her collaborators have found a number of practical issues with ADVI. It has to solve an optimization problem and can do so only approximately. So, ADVI can still require a lot of computation time and user effort to determine whether the approximate solution is good enough. And once it arrives at a solution, it tends to provide poor uncertainty estimates.
Rather than reinventing the wheel, the team took many ideas from ADVI but turned them around to create a technique called deterministic ADVI (DADVI) that doesn’t have these downsides.
With DADVI, it is very clear when the optimization is finished, so a user won’t need to spend extra computation time to ensure that the best solution has been found. DADVI also permits the incorporation of more powerful optimization methods that give it an additional speed and performance boost.
Once it reaches a result, DADVI is set up to allow the use of uncertainty corrections. These corrections make its uncertainty estimates much more accurate than those of ADVI.
DADVI also enables the user to clearly see how much error they have incurred in the approximation to the optimization problem. This prevents a user from needlessly running the optimization again and again with more and more resources to try and reduce the error.

“We wanted to see if we could live up to the promise of black-box inference in the sense of, once the user makes their model, they can just run Bayesian inference and don’t have to derive everything by hand, they don’t need to figure out when to stop their algorithm, and they have a sense of how accurate their approximate solution is,” Broderick says.
Defying conventional wisdom
DADVI can be more effective than ADVI because it uses an efficient approximation method, called sample average approximation, which estimates an unknown quantity by taking a series of exact steps.
Because the steps along the way are exact, it is clear when the objective has been reached. Plus, getting to that objective typically requires fewer steps.
Often, researchers expect sample average approximation to be more computationally intensive than a more popular method, known as stochastic gradient, which is used by ADVI. But Broderick and her collaborators showed that, in many applications, this is not the case.
“A lot of problems really do have special structure, and you can be so much more efficient and get better performance by taking advantage of that special structure. That is something we have really seen in this paper,” she adds.
They tested DADVI on a number of real-world models and datasets, including a model used by economists to evaluate the effectiveness of microcredit loans and one used in ecology to determine whether a species is present at a particular site.
Across the board, they found that DADVI can estimate unknown parameters faster and more reliably than other methods, and achieves as good or better accuracy than ADVI. Because it is easier to use than other techniques, DADVI could offer a boost to scientists in a wide variety of fields.
In the future, the researchers want to dig deeper into correction methods for uncertainty estimates so they can better understand why these corrections can produce such accurate uncertainties, and when they could fall short.
This research was supported by a National Science Foundation CAREER Award and the U.S. Office of Naval Research. More

The electron is the basic unit of electricity, as it carries a single negative charge. This is what we’re taught in high school physics, and it is overwhelmingly the case in most materials in nature.
But in very special states of matter, electrons can splinter into fractions of their whole. This phenomenon, known as “fractional charge,” is exceedingly rare, and if it can be corralled and controlled, the exotic electronic state could help to build resilient, fault-tolerant quantum computers.
To date, this effect, known to physicists as the “fractional quantum Hall effect,” has been observed a handful of times, and mostly under very high, carefully maintained magnetic fields. Only recently have scientists seen the effect in a material that did not require such powerful magnetic manipulation.
Now, MIT physicists have observed the elusive fractional charge effect, this time in a simpler material: five layers of graphene — an atom-thin layer of carbon that stems from graphite and common pencil lead. They report their results in Nature.
They found that when five sheets of graphene are stacked like steps on a staircase, the resulting structure inherently provides just the right conditions for electrons to pass through as fractions of their total charge, with no need for any external magnetic field.
The results are the first evidence of the “fractional quantum anomalous Hall effect” (the term “anomalous” refers to the absence of a magnetic field) in crystalline graphene, a material that physicists did not expect to exhibit this effect.
“This five-layer graphene is a material system where many good surprises happen,” says study author Long Ju, assistant professor of physics at MIT. “Fractional charge is just so exotic, and now we can realize this effect with a much simpler system and without a magnetic field. That in itself is important for fundamental physics. And it could enable the possibility for a type of quantum computing that is more robust against perturbation.”
Ju’s MIT co-authors are lead author Zhengguang Lu, Tonghang Han, Yuxuan Yao, Aidan Reddy, Jixiang Yang, Junseok Seo, and Liang Fu, along with Kenji Watanabe and Takashi Taniguchi at the National Institute for Materials Science in Japan.

A bizarre state
The fractional quantum Hall effect is an example of the weird phenomena that can arise when particles shift from behaving as individual units to acting together as a whole. This collective “correlated” behavior emerges in special states, for instance when electrons are slowed from their normally frenetic pace to a crawl that enables the particles to sense each other and interact. These interactions can produce rare electronic states, such as the seemingly unorthodox splitting of an electron’s charge.
In 1982, scientists discovered the fractional quantum Hall effect in heterostructures of gallium arsenide, where a gas of electrons confined in a two-dimensional plane is placed under high magnetic fields. The discovery later won the group a Nobel Prize in Physics.
“[The discovery] was a very big deal, because these unit charges interacting in a way to give something like fractional charge was very, very bizarre,” Ju says. “At the time, there were no theory predictions, and the experiments surprised everyone.”
Those researchers achieved their groundbreaking results using magnetic fields to slow down the material’s electrons enough for them to interact. The fields they worked with were about 10 times stronger than what typically powers an MRI machine.
In August 2023, scientists at the University of Washington reported the first evidence of fractional charge without a magnetic field. They observed this “anomalous” version of the effect, in a twisted semiconductor called molybdenum ditelluride. The group prepared the material in a specific configuration, which theorists predicted would give the material an inherent magnetic field, enough to encourage electrons to fractionalize without any external magnetic control.

The “no magnets” result opened a promising route to topological quantum computing — a more secure form of quantum computing, in which the added ingredient of topology (a property that remains unchanged in the face of weak deformation or disturbance) gives a qubit added protection when carrying out a computation. This computation scheme is based on a combination of fractional quantum Hall effect and a superconductor. It used to be almost impossible to realize: One needs a strong magnetic field to get fractional charge, while the same magnetic field will usually kill the superconductor. In this case the fractional charges would serve as a qubit (the basic unit of a quantum computer).
Making steps
That same month, Ju and his team happened to also observe signs of anomalous fractional charge in graphene — a material for which there had been no predictions for exhibiting such an effect.
Ju’s group has been exploring electronic behavior in graphene, which by itself has exhibited exceptional properties. Most recently, Ju’s group has looked into pentalayer graphene — a structure of five graphene sheets, each stacked slightly off from the other, like steps on a staircase. Such pentalayer graphene structure is embedded in graphite and can be obtained by exfoliation using Scotch tape. When placed in a refrigerator at ultracold temperatures, the structure’s electrons slow to a crawl and interact in ways they normally wouldn’t when whizzing around at higher temperatures.
In their new work, the researchers did some calculations and found that electrons might interact with each other even more strongly if the pentalayer structure were aligned with hexagonal boron nitride (hBN) — a material that has a similar atomic structure to that of graphene, but with slightly different dimensions. In combination, the two materials should produce a moiré superlattice — an intricate, scaffold-like atomic structure that could slow electrons down in ways that mimic a magnetic field.
“We did these calculations, then thought, let’s go for it,” says Ju, who happened to install a new dilution refrigerator in his MIT lab last summer, which the team planned to use to cool materials down to ultralow temperatures, to study exotic electronic behavior.
The researchers fabricated two samples of the hybrid graphene structure by first exfoliating graphene layers from a block of graphite, then using optical tools to identify five-layered flakes in the steplike configuration. They then stamped the graphene flake onto an hBN flake and placed a second hBN flake over the graphene structure. Finally, they attached electrodes to the structure and placed it in the refrigerator, set to near absolute zero.
As they applied a current to the material and measured the voltage output, they started to see signatures of fractional charge, where the voltage equals the current multiplied by a fractional number and some fundamental physics constants.
“The day we saw it, we didn’t recognize it at first,” says first author Lu. “Then we started to shout as we realized, this was really big. It was a completely surprising moment.”
“This was probably the first serious samples we put in the new fridge,” adds co-first author Han. “Once we calmed down, we looked in detail to make sure that what we were seeing was real.”
With further analysis, the team confirmed that the graphene structure indeed exhibited the fractional quantum anomalous Hall effect. It is the first time the effect has been seen in graphene.
“Graphene can also be a superconductor,” Ju says. “So, you could have two totally different effects in the same material, right next to each other. If you use graphene to talk to graphene, it avoids a lot of unwanted effects when bridging graphene with other materials.”
For now, the group is continuing to explore multilayer graphene for other rare electronic states.
“We are diving in to explore many fundamental physics ideas and applications,” he says. “We know there will be more to come.”
This research is supported in part by the Sloan Foundation, and the National Science Foundation. More

In the blink of an eye, the unruly, superheated plasma that drives a fusion reaction can lose its stability and escape the strong magnetic fields confining it within the donut-shaped fusion reactor. These getaways frequently spell the end of the reaction, posing a core challenge to developing fusion as a non-polluting, virtually limitless energy source.
But a Princeton-led team composed of engineers, physicists, and data scientists from the University and the Princeton Plasma Physics Laboratory (PPPL) have harnessed the power of artificial intelligence to predict — and then avoid — the formation of a specific plasma problem in real time.
In experiments at the DIII-D National Fusion Facility in San Diego, the researchers demonstrated their model, trained only on past experimental data, could forecast potential plasma instabilities known as tearing mode instabilities up to 300 milliseconds in advance. While that leaves no more than enough time for a slow blink in humans, it was plenty of time for the AI controller to change certain operating parameters to avoid what would have developed into a tear within the plasma’s magnetic field lines, upsetting its equilibrium and opening the door for a reaction-ending escape.
“By learning from past experiments, rather than incorporating information from physics-based models, the AI could develop a final control policy that supported a stable, high-powered plasma regime in real time, at a real reactor,” said research leader Egemen Kolemen, associate professor of mechanical and aerospace engineering and the Andlinger Center for Energy and the Environment, as well as staff research physicist at PPPL.
The research opens the door for more dynamic control of a fusion reaction than current approaches, and it provides a foundation for using artificial intelligence to solve a broad range of plasma instabilities, which have long been obstacles to achieving a sustained fusion reaction. The team published their findings in Nature on February 21.
“Previous studies have generally focused on either suppressing or mitigating the effects of these tearing instabilities after they occur in the plasma,” said first author Jaemin Seo, an assistant professor of physics at Chung-Ang University in South Korea who performed much of the work while a postdoctoral researcher in Kolemen’s group. “But our approach allows us to predict and avoid those instabilities before they ever appear.”
Superheated plasma swirling in a donut-shaped device
Fusion takes place when two atoms — usually light atoms like hydrogen — come together to form one heavier atom, releasing a large amount of energy in the process. The process powers the Sun, and, by extension, makes life on Earth possible.

However, getting the two atoms to fuse is tricky, as it takes massive amounts of pressure and energy for the two atoms to overcome their mutual repulsion.
Fortunately for the Sun, its massive gravitational pull and extremely high pressures at its core allow fusion reactions to proceed. To replicate a similar process on the Earth, scientists instead use extremely hot plasma and extremely strong magnets.
In donut-shaped devices known as tokamaks — sometimes referred to as “stars in jars” — magnetic fields struggle to contain plasmas that reach above 100 million degrees Celsius, hotter than the center of the Sun.
While there are many types of plasma instabilities that can terminate the reaction, the Princeton team concentrated on solving tearing mode instabilities, a disturbance in which the magnetic field lines within a plasma actually break and create an opportunity for the plasma’s subsequent escape.
“Tearing mode instabilities are one of the major causes of plasma disruption, and they will become even more prominent as we try to run fusion reactions at the high powers required to produce enough energy,” said Seo. “They are an important challenge for us to solve.”
Fusing artificial intelligence and plasma physics
Since tearing mode instabilities can form and derail a fusion reaction in milliseconds, the researchers turned to artificial intelligence for its ability to quickly process and act in response to new data.

But the process to develop an effective AI controller was not as simple as trying out a few things on a tokamak, where time is limited, and the stakes are high.
Co-author Azarakhsh Jalalvand, a research scholar in Kolemen’s group, compared teaching an algorithm to run a fusion reaction in a tokamak to teaching someone how to fly a plane.
“You wouldn’t teach someone by handing them a set of keys and telling them to try their best,” Jalalvand said. “Instead, you’d have them practice on a very intricate flight simulator until they’ve learned enough to try out the real thing.”
Like developing a flight simulator, the Princeton team used data from past experiments at the DIII-D tokamak to construct a deep neural network capable of predicting the likelihood of a future tearing instability based on real-time plasma characteristics.
They used that neural network to train a reinforcement learning algorithm. Like a pilot trainee, the reinforcement learning algorithm could try out different strategies for controlling plasma, learning through trial and error which strategies worked and which did not within the safety of a simulated environment.
“We don’t teach the reinforcement learning model all of the complex physics of a fusion reaction,” Jalalvand said. “We tell it what the goal is — to maintain a high-powered reaction — what to avoid — a tearing mode instability — and the knobs it can turn to achieve those outcomes. Over time, it learns the optimal pathway for achieving the goal of high power while avoiding the punishment of an instability.”
While the model went through countless simulated fusion experiments, trying to find ways to maintain high power levels while avoiding instabilities, co-author SangKyeun Kim could observe and refine its actions.
“In the background, we can see the intentions of the model,” said Kim, a staff research scientist at PPPL and former postdoctoral researcher in Kolemen’s group. “Some of the chnges that the model wants are too rapid, so we work to smooth and calm the model. As humans, we arbitrate between what the AI wants to do and what the tokamak can accommodate.”
Once they were confident in the AI controller’s abilities, they tested it during an actual fusion experiment at the D-III D tokamak, observing as the controller made real-time changes to certain tokamak parameters to avoid the onset of an instability. These parameters included changing the shape of the plasma and the strength of the beams inputting power into the reaction.
“Being able to predict instabilities ahead of time can make it easier to run these reactions than current approaches, which are more passive,” said Kim. “We no longer have to wait for the instabilities to occur and then take quick corrective action before the plasma becomes disrupted.”
Powering into the future
While the researchers said the work is a promising proof-of-concept demonstrating how artificial intelligence can effectively control fusion reactions, it is only one of many next steps already ongoing in Kolemen’s group to advance the field of fusion research.
The first step is to get more evidence of the AI controller in action at the DIII-D tokamak, and then expand the controller to function at other tokamaks.
“We have strong evidence that the controller works quite well at DIII-D, but we need more data to show that it can work in a number of different situations,” said first author Seo. “We want to work toward something more universal.”
A second line of research involves expanding the algorithm to handle many different control problems at the same time. While the current model uses a limited number of diagnostics to avoid one specific type of instability, the researchers could provide data on other types of instabilities and give access to more knobs for the AI controller to tune.
“You could imagine one large reward function that turns many different knobs to simultaneously control for several types of instabilities,” said co-author Ricardo Shousha, a postdoc at PPPL and former graduate student in Kolemen’s group who provided support for the experiments at DIII-D.
And on the route to developing better AI controllers for fusion reactions, researchers might also gain more understanding of the underlying physics. By studying the AI controller’s decisions as it attempts to contain the plasma, which can be radically different than what traditional approaches might prescribe, artificial intelligence may be not only a tool to control fusion reactions but also a teaching resource.
“Eventually, it may be more than just a one-way interaction of scientists developing and deploying these AI models,” said Kolemen. “By studying them in more detail, they may have certain things that they can teach us too.”
The work was supported by the U.S. Department of Energy’s Office of Fusion Energy Sciences, as well as the National Research Foundation of Korea (NRF). The authors also acknowledge the use of the DIII-D National Fusion Facility, a Department of Energy Office of Science user facility. More