More stories

  • in

    New AI algorithm may improve autoimmune disease prediction and therapies

    A new advanced artificial intelligence (AI) algorithm may lead to better — and earlier — predictions and novel therapies for autoimmune diseases, which involve the immune system mistakenly attacking their body’s own healthy cells and tissues. The algorithm digs into the genetic code underlying the conditions to more accurately model how genes associated with specific autoimmune diseases are expressed and regulated and to identify additional genes of risk.
    The work, developed by a team led by Penn State College of Medicine researchers, outperforms existing methodologies and identified 26% more novel gene and trait associations, the researchers said. They published their work today (May 20) in Nature Communications.
    “We all carry some DNA mutations, and we need to figure out how any one of these mutations may influence gene expression linked to disease so we can predict disease risk early. This is especially important for autoimmune disease,” said Dajiang Liu, distinguished professor, vice chair for research, and director of artificial intelligence and biomedical informatics at the Penn State College of Medicine and co-senior author of the study. “If an AI algorithm can more accurately predict disease risk, it means we can carry out interventions earlier.”
    Genetics often underpin disease development. Variations in DNA can influence gene expression, or the process by which the information in DNA is converted into functional products like a protein. How much or how little a gene is expressed can influence disease risk.
    Genome-wide association studies (GWAS), a popular approach in human genetics research, can home in on regions of the genome associated with a particular disease or trait but can’t pinpoint the specific genes that affect disease risks. It’s like sharing your location with a friend with the precise location setting turned off on your smartphone — the city might be obvious, but the address is obscured. Existing methods are also limited in the granularity of its analysis. Gene expression can be specific to certain types of cells. If the analysis doesn’t distinguish between distinct cell types, the results may overlook real causal relationships between genetic variants and gene expression.
    The research team’s method, dubbed EXPRESSO for EXpression PREdiction with Summary Statistics Only, applies a more advanced artificial intelligence algorithm and analyzes data from single-cell expression quantitative trait loci, a type of data that links genetic variants to the genes they regulate. It also integrates 3D genomic data and epigenetics — which measures how genes may be modified by environment to influence disease — into its modeling. The team applied EXPRESSO to GWAS datasets for 14 autoimmune diseases, including lupus, Crohn’s disease, ulcerative colitis and rheumatoid arthritis.
    “With this new method, we were able to identify many more risk genes for autoimmune disease that actually have cell-type specific effects, meaning that they only have effects in a particular cell type and not others,” said Bibo Jiang, assistant professor at the Penn State College of Medicine and senior author of the study.

    The team then used this information to identify potential therapeutics for autoimmune disease. Currently, there aren’t good long-term treatment options, they said.
    “Most treatments are designed to mitigate symptoms, not cure the disease. It’s a dilemma knowing that autoimmune disease needs long-term treatment, but the existing treatments often have such bad side effects that they can’t be used for long. Yet, genomics and AI offer a promising route to develop novel therapeutics,” said Laura Carrel, professor of biochemistry and molecular biology at the Penn State College of Medicine and co-senior author of the study.
    The team’s work pointed to drug compounds that could reverse gene expression in cell types associated with an autoimmune disease, such as vitamin K for ulcerative colitis and metformin, which is typically prescribed for type 2 diabetes, for type 1 diabetes. These drugs, already approved by the Food and Drug Administration as safe and effective for treating other diseases, could potentially be repurposed.
    The research team is working with collaborators to validate their findings in a laboratory setting and, ultimately, in clinical trials.
    Lida Wang, a doctoral student in the biostatistics program, and Chachrit Khunsriraksakul, who earned a doctorate in bioinformatics and geonomics in 2022 and his medical degree in May from Penn State, co-led the study. Other Penn State College of Medicine authors on the paper include: Havell Markus, who is pursuing a doctorate and a medical degree; Dieyi Chen, doctoral candidate; Fan Zhang, graduate student; and Fang Chen, postdoctoral scholar. Xiaowei Zhan, associate professor at UT Southwestern Medical Center, also contributed to the paper.
    Funding from the National Institutes of Health (grant numbers R01HG011035, R01AI174108 and R01ES036042) and the Artificial Intelligence and Biomedical Informatics pilot grant from the Penn State College of Medicine supported this work. More

  • in

    Physicists propose path to faster, more flexible robots

    In a May 15 paper released in the journal Physical Review Letters, Virginia Tech physicists revealed a microscopic phenomenon that could greatly improve the performance of soft devices, such as agile flexible robots or microscopic capsules for drug delivery.
    The paper, written by doctoral candidate Chinmay Katke, assistant professor C. Nadir Kaplan, and co-author Peter A. Korevaar from Radboud University in the Netherlands, proposes a new physical mechanism that could speed up the expansion and contraction of hydrogels. For one thing, this opens up the possibility for hydrogels to replace rubber-based materials used to make flexible robots — enabling these fabricated materials to perhaps move with a speed and dexterity close to that of human hands.
    Soft robots are already being used in manufacturing, where a hand-like device is programmed to grab an item from a conveyer belt — picture a hot dog or piece of soap — and place it in a container to be packaged. But the ones in use now lean on hydraulics or pneumatics to change the shape of the “hand” to pick up the item.
    Akin to our own body, hydrogels mostly contain water and are everywhere around us, e.g., food jelly and shaving gel. Katke, Korevaar, and Kaplan’s research appears to have found a method that allows hydrogels to swell and contract much more quickly, which would improve their flexibility and capability to function in different settings.
    Living organisms use osmosis for such activities as bursting seed dispersing fruits in plants or absorbing water in the intestine. Normally, we think of osmosis as a flow of water moving through a membrane, with bigger molecules like polymers unable to move through. Such membranes are called semi-permeable membranes and were thought to be necessary to trigger osmosis.
    Previously, Korevaar and Kaplan had done experiments by using a thin layer of hydrogel film comprised of polyacrylic acid. They had observed that even though the hydrogel film allows both water and ions to pass through and is not selective, the hydrogel rapidly swells due to osmosis when ions are released inside the hydrogel and shrinks back again.
    Katke, Korevaar, and Kaplan developed a new theory to explain the above observation. This theory tells that microscopic interactions between ions and polyacrylic acid can make hydrogel swell when the released ions inside the hydrogel are unevenly spread out. They called this “diffusio-phoretic swelling of the hydrogels.” Furthermore, this newly discovered mechanism allows hydrogels to swell much faster than what has been previously possible.

    Why is that change important?
    Kaplan explained: Soft agile robots are currently made with rubber, which “does the job but their shapes are changed hydraulically or pneumatically. This is not desired because it is difficult to imprint a network of tubes into these robots to deliver air or fluid into them.”
    Imagine, Kaplan said, how many different things you can do with your hand and how fast you can do them owing to your neural network and the motion of ions under your skin. Because the rubber and hydraulics are not as versatile as your biological tissues, which is a hydrogel, state-of-the-art soft robots can only do a limited number of movements.”
    Katke explained that the process they have researched allows the hydrogels to change shape then change back to their original form “significantly faster this way” in soft robots that are larger than ever before.
    At present, only microscopic-sized hydrogel robots can respond to a chemical signal quickly enough to be useful and larger ones require hours to change shape, Katke said. By using the new diffusio-phoresis method, soft robots as large as a centimeter may be able to transform in just a few seconds, which is subject to further studies.
    Larger agile soft robots that could respond quickly could improve assistive devices in healthcare, “pick-and-place” functions in manufacturing, search and rescue operations, cosmetics used for skincare, and contact lenses. More

  • in

    World’s smallest quantum light detector on a silicon chip

    Researchers at the University of Bristol have made an important breakthrough in scaling quantum technology by integrating the world’s tiniest quantum light detector onto a silicon chip.
    A critical moment in unlocking the information age was when scientists and engineers were first able to miniaturise transistors onto cheap micro-chips in the 1960s.
    Now, for the first time, University of Bristol academics have demonstrated the integration of a quantum light detector — smaller than a human hair — onto a silicon chip, moving us one step closer to the age of quantum technologies using light.
    Making high performance electronics and photonics at scale is fundamental to realizing the next generation of advanced information technologies. Figuring out how to make quantum technologies in existing commercial facilities is an ongoing international effort being tackled by university research and companies around the world.
    It could prove crucial for quantum computing to be able to make high performance quantum hardware at scale due to the vast amount of components anticipated to build even a single machine.
    In pursuit of this goal, researchers at the University of Bristol have demonstrated a type of quantum light detector that is implemented on a chip with a circuit that occupies 80 micrometers by 220 micrometers.
    Critically, the small size means the quantum light detector can be fast, which is key to unlocking high speed quantum communications and enabling high speed operation of optical quantum computers.

    The use of established and commercially accessible fabrication techniques helps the prospects for early incorporation into other technologies such as sensing and communications.
    “These types of detectors are called homodyne detectors, and they pop up everywhere in applications across quantum optics” explains Professor Jonathan Matthews, who led the research and is Director of the Quantum Engineering Technology Labs. “They operate at room temperature, and you can use them for quantum communications, in incredibly sensitive sensors — like state-of-the-art gravitational wave detectors — and there are designs of quantum computers that would use these detectors.”
    In 2021 the Bristol team showed how linking a photonics chip with a separate electronics chip can increase speed of quantum light detectors — now with a single electronic-photonic integrated chip, the team have further increased speed by a factor of 10 whilst reducing footprint by a factor of 50.
    While these detectors are fast and small, they are also sensitive.
    “The key to measuring quantum light is sensitivity to quantum noise” explains author Dr Giacomo Ferranti. “Quantum mechanics is responsible for a minute, fundamental level of noise in all optical systems. The behaviour of this noise reveals information about what kind of quantum light is travelling in the system, it can determine how sensitive an optical sensor can be, and it can be used to mathematically reconstruct quantum states. In our study it was important to show that making the detector smaller and faster did not block its sensitivity for measuring quantum states.”
    The authors note that there is more exciting research to do in integrating other disruptive quantum technology hardware down to the chip scale. With the new detector, the efficiency needs to improve, and there is work to be done to trial the detector in lots of different applications.
    Professor Matthews added: “We built the detector with a commercially accessible foundry in order to make its applications more accessible. While we are incredibly excited by the implications across a range of quantum technology, it is critical that we as a community continue to tackle the challenge of scalable fabrication of quantum technology. Without demonstrating truly scalable fabrication of quantum hardware, the impact and benefits of quantum technology will be delayed and limited.” More

  • in

    Diamond glitter: A play of colors with artificial DNA crystals

    Using DNA origami, LMU researchers have built a diamond lattice with a periodicity of hundreds of nanometers — a new approach for manufacturing semiconductors for visible light.
    The shimmering of butterfly wings in bright colors does not emerge from pigments. Rather, it is photonic crystals that are responsible for the play of colors. Their periodic nanostructure allows light at certain wavelengths to pass through while reflecting other wavelengths. This causes the wing scales, which are in fact transparent, to appear so magnificently colored. For research teams, the manufacture of artificial photonic crystals for visible light wavelengths has been a major challenge and motivation ever since they were predicted by theorists more than 35 years ago. “Photonic crystals have a versatile range of applications. They have been employed to develop more efficient solar cells, innovative optical waveguides, and materials for quantum communication. However, they have been very laborious to manufacture,” explains Dr. Gregor Posnjak. The physicist is a postdoc in the research group of LMU Professor Tim Liedl, whose work is funded by the “e-conversion” Cluster of Excellence and the European Research Council. Using DNA nanotechnology, the team has developed a new approach for the manufacture of photonic crystals. Their results have now been published in the journal Science.
    Diamond structure out of strands of DNA
    In contrast to lithographic techniques, the LMU team uses a method called DNA origami to design and synthesize building blocks, which then self-assemble into a specific lattice structure. “It’s long been known that the diamond lattice theoretically has an optimal geometry for photonic crystals. In diamonds, each carbon atom is bonded to four other carbon atoms. Our challenge consisted in enlarging the structure of a diamond crystal by a factor of 500, so that the spaces between the building blocks correspond with the wavelength of light,” explains Tim Liedl. “We increased the periodicity of the lattice to 170 nanometers by replacing the individual atoms with larger building blocks — in our case, through DNA origami,” says Posnjak.
    The perfect molecule folding technique
    What sounds like magic is actually a specialty of the Liedl group, one of the world’s leading research teams in DNA origami and self-assembly. For this purpose, the scientists use a long, ring-shaped DNA strand (consisting of around 8,000 bases) and a set of 200 short DNA staples. “The latter control the folding of the longer DNA strand into virtually any shape at all — akin to origami masters, who fold pieces of paper into intricate objects. As such, the clamps are a means of determining how the DNA origami objects combine to form the desired diamond lattice,” says the LMU postdoctoral researcher. The DNA origami building blocks form crystals of approximately ten micrometers in size, which are deposited on a substrate and then passed on to a cooperating research group from the Walter Schottky Institute at the Technical University of Munich (TUM): The team led by Professor Ian Sharp (also funded by the “e-conversion” Cluster of Excellence) is able to deposit individual atomic layers of titanium dioxide on all surfaces of the DNA origami crystals. “The DNA origami diamond lattice serves as scaffolding for titanium dioxide, which, on account of its high index of refraction, determines the photonic properties of the lattice. After coating, our photonic crystal does not allow UV light with a wavelength of about 300 nanometers to pass through, but rather reflects it,” explains Posnjak. The wavelength of the reflected light can be controlled via the thickness of the titanium dioxide layer.
    DNA origami could boost photonics
    For photonic crystals that work in the infrared range, classic lithographic techniques are suitable but laborious and expensive. In the wavelength range of visible and UV light, lithographic methods have not been successful to date. “Consequently, the comparatively easy manufacturing process using the self-assembly of DNA origami in an aqueous solution offers a powerful alternative for producing structures in the desired size cost-effectively and in larger quantities,” says Prof. Tim Liedl. He is convinced that the unique structure with its large pores, which are chemically addressable, will stimulate further research — for example, in the domain of energy harvesting and storage. In the same issue of Science, a collaboration led by prof. Petr Šulc of Arizona State University and TUM presents a theoretical framework for designing diverse crystalline lattices from patchy colloids, and experimentally demonstrates the method by utilizing DNA origami building blocks to form a pyrochlore lattice, which potentially also could be used for photonic applications. More

  • in

    New AI tool to help beat brain tumors

    A new AI tool to more quickly and accurately classify brain tumours has been developed by researchers at The Australian National University (ANU).
    According to Dr Danh-Tai Hoang, precision in diagnosing and categorising tumours is crucial for effective patient treatment.
    “The current gold standard for identifying different kinds of brain tumours is DNA methylation-based profiling,” Dr Hoang said.
    “DNA methylation acts like a switch to control gene activity, and which genes are turned on or off.
    “But the time it takes to do this kind of testing can be a major drawback, often requiring several weeks or more when patients might be relying on quick decisions on therapies.
    “There’s also a lack of availability of these tests in nearly all hospitals worldwide.”
    To address these challenges, the ANU researchers, in collaboration with experts from the National Cancer Institute in the United States (US), developed DEPLOY, a way to predict DNA methylation and subsequently classify brain tumours into 10 major subtypes.

    DEPLOY draws on microscopic pictures of a patient’s tissue called histopathology images.
    The model was trained and validated on large datasets of approximately 4,000 patients from across the US and Europe.
    “Remarkably, DEPLOY achieved an unprecedented accuracy of 95 per cent,” Dr Hoang said.
    “Furthermore, when given a subset of 309 particularly difficult to classify samples, DEPLOY was able to provide a diagnosis that was more clinically relevant than what was initially provided by pathologists.
    “This shows the potential future role of DEPLOY as a complementary tool, adding to a pathologist’s initial diagnosis, or even prompting re-evaluation in the case of disparities.”
    The researchers believe DEPLOY could eventually be used to help classify other types of cancer as well. More

  • in

    A powerful tool speeds success in achieving highly efficient thermoelectric materials

    Thermoelectric materials could play an important role in the clean energy transition, as they can produce electricity from sources of heat that would otherwise go to waste without generating additional greenhouse gases or requiring large up-front investment. But their promise has been slowed by the fact that most current thermoelectric materials don’t efficiently produce enough power to be useful for many practical applications.
    The search for new, more efficient materials involving complex chemical compositions has been labor-intensive, requiring experimental testing of each proposed new multi-material composition, and has often involved the use of toxic or rare elements. In a paper published Thursday, May 16, in the journal Science, researchers from the University of Houston and Rice University report a new approach to predict the realization of band convergence in a series of materials and, after demonstrating that one so-designed material, a p-type Zintl compound, would offer highly efficient thermoelectric performance, fabricated a thermoelectric module. They reported a heat-to-electricity conversion efficiency exceeding 10% at a temperature difference of 475 kelvin, or about 855 degrees Fahrenheit.
    Zhifeng Ren, director of the Texas Center for Superconductivity at UH (TcSUH) and corresponding author for the paper, said the materials’ performance remained stable for more than two years.
    While a variety of approaches have been used to improve efficiency, a concept known as electronic band convergence has gained attention for its potential to improve thermoelectric performance. “It is normally difficult to get high performance from thermoelectric materials because not all of the electronic bands in a material contribute,” Ren said. “It’s even more difficult to make a complex material where all of the bands work at the same time in order to get the best performance.”
    For this work, he said, the scientists first focused on devising a calculation to determine how to build a material in which all the different energy bands can contribute to the overall performance. They then demonstrated that the calculation worked in practice as well as in theory, building a module to further verify the obtained high performance at the device level.
    Band convergence is considered a good approach for improving thermoelectric materials because it increases the thermoelectric power factor, which is related to the actual output power of the thermoelectric module. But until now, discovering new materials with strong band convergence was time-consuming and resulted in many false starts. “The standard approach is trial and error,” said Ren, who is also the Paul C.W. Chu and May P. Chern Endowed Chair in Condensed Matter Physics at UH. “Instead of doing a lot of experiments, this method allows us to eliminate unnecessary possibilities that won’t give better results.”
    To efficiently predict how to create the most effective material, the researchers used a high-entropy Zintl alloy, YbxCa1-xMgyZn2-ySb2, as a case study, designing a series of compositions through which band convergence was achieved simultaneously in all of the compositions.

    Ren described how it works like this: If a team of 10 people try to lift an object, the taller members will carry most of the load while the shorter members do not contribute as much. In band convergence, the goal is to make all the band team members more similar — tall band members would be shorter, in this example, and short members taller — so all can contribute to carrying the overall load.
    Here, the researchers started with four parent compounds containing five elements in total — ytterbium, calcium, magnesium, zinc and antimony — running calculations to determine which combinations of the parent compounds could reach band convergence. Once that was determined, they chose the best among these high-performance compositions to construct the thermoelectric device.
    “Without this method, you would have to experiment and try all possibilities,” said Xin Shi, a UH graduate student in Ren’s group and lead author on the paper. “There’s no other way you can do that. Now, we do a calculation first, we design a material, and then make it and test it.”
    The calculation method could be used for other multi-compound materials, too, allowing researchers to use this approach to create new thermoelectric materials. Once the proper parent compounds are identified, the calculation determines what ratio of each should be used in the final alloy.
    In addition to Ren and Shi, the paper’s authors include Dr. Shaowei Song, a researcher at the Texas Center for Superconductivity, and Dr. Guanhui Gao from the Department of Materials Science and Nanoengineering at Rice. Gao is now at UH. More

  • in

    Large language models can’t effectively recognize users’ motivation, but can support behavior change for those ready to act

    Large language model-based chatbots have the potential to promote healthy changes in behavior. But researchers from the ACTION Lab at the University of Illinois Urbana-Champaign have found that the artificial intelligence tools don’t effectively recognize certain motivational states of users and therefore don’t provide them with appropriate information.
    Michelle Bak, a doctoral student in information sciences, and information sciences professor Jessie Chin reported their research in the Journal of the American Medical Informatics Association.
    Large language model-based chatbots — also known as generative conversational agents — have been used increasingly in healthcare for patient education, assessment and management. Bak and Chin wanted to know if they also could be useful for promoting behavior change.
    Chin said previous studies showed that existing algorithms did not accurately identify various stages of users’ motivation. She and Bak designed a study to test how well large language models, which are used to train chatbots, identify motivational states and provide appropriate information to support behavior change.
    They evaluated large language models from ChatGPT, Google Bard and Llama 2 on a series of 25 different scenarios they designed that targeted health needs that included low physical activity, diet and nutrition concerns, mental health challenges, cancer screening and diagnosis, and others such as sexually transmitted disease and substance dependency.
    In the scenarios, the researchers used each of the five motivational stages of behavior change: resistance to change and lacking awareness of problem behavior; increased awareness of problem behavior but ambivalent about making changes; intention to take action with small steps toward change; initiation of behavior change with a commitment to maintain it; and successfully sustaining the behavior change for six months with a commitment to maintain it.
    The study found that large language models can identify motivational states and provide relevant information when a user has established goals and a commitment to take action. However, in the initial stages when users are hesitant or ambivalent about behavior change, the chatbot is unable to recognize those motivational states and provide appropriate information to guide them to the next stage of change.

    Chin said that language models don’t detect motivation well because they are trained to represent the relevance of a user’s language, but they don’t understand the difference between a user who is thinking about a change but is still hesitant and a user who has the intention to take action. Additionally, she said, the way users generate queries is not semantically different for the different stages of motivation, so it’s not obvious from the language what their motivational states are.
    “Once a person knows they want to start changing their behavior, large language models can provide the right information. But if they say, ‘I’m thinking about a change. I have intentions but I’m not ready to start action,’ that is the state where large language models can’t understand the difference,” Chin said.
    The study results found that when people were resistant to habit change, the large language models failed to provide information to help them evaluate their problem behavior and its causes and consequences and assess how their environment influenced the behavior. For example, if someone is resistant to increasing their level of physical activity, providing information to help them evaluate the negative consequences of sedentary lifestyles is more likely to be effective in motivating users through emotional engagement than information about joining a gym. Without information that engaged with the users’ motivations, the language models failed to generate a sense of readiness and the emotional impetus to progress with behavior change, Bak and Chin reported.
    Once a user decided to take action, the large language models provided adequate information to help them move toward their goals. Those who had already taken steps to change their behaviors received information about replacing problem behaviors with desired health behaviors and seeking support from others, the study found.
    However, the large language models didn’t provide information to those users who were already working to change their behaviors about using a reward system to maintain motivation or about reducing the stimuli in their environment that might increase the risk of a relapse of the problem behavior, the researchers found.
    “The large language model-based chatbots provide resources on getting external help, such as social support. They’re lacking information on how to control the environment to eliminate a stimulus that reinforces problem behavior,” Bak said.
    Large language models “are not ready to recognize the motivation states from natural language conversations, but have the potential to provide support on behavior change when people have strong motivations and readiness to take actions,” the researchers wrote.
    Chin said future studies will consider how to finetune large language models to use linguistic cues, information search patterns and social determinants of health to better understand a users’ motivational states, as well as providing the models with more specific knowledge for helping people change their behaviors. More

  • in

    Scientists use generative AI to answer complex questions in physics

    When water freezes, it transitions from a liquid phase to a solid phase, resulting in a drastic change in properties like density and volume. Phase transitions in water are so common most of us probably don’t even think about them, but phase transitions in novel materials or complex physical systems are an important area of study.
    To fully understand these systems, scientists must be able to recognize phases and detect the transitions between. But how to quantify phase changes in an unknown system is often unclear, especially when data are scarce.
    Researchers from MIT and the University of Basel in Switzerland applied generative artificial intelligence models to this problem, developing a new machine-learning framework that can automatically map out phase diagrams for novel physical systems.
    Their physics-informed machine-learning approach is more efficient than laborious, manual techniques which rely on theoretical expertise. Importantly, because their approach leverages generative models, it does not require huge, labeled training datasets used in other machine-learning techniques.
    Such a framework could help scientists investigate the thermodynamic properties of novel materials or detect entanglement in quantum systems, for instance. Ultimately, this technique could make it possible for scientists to discover unknown phases of matter autonomously.
    “If you have a new system with fully unknown properties, how would you choose which observable quantity to study? The hope, at least with data-driven tools, is that you could scan large new systems in an automated way, and it will point you to important changes in the system. This might be a tool in the pipeline of automated scientific discovery of new, exotic properties of phases,” says Frank Schäfer, a postdoc in the Julia Lab in the Computer Science and Artificial Intelligence Laboratory (CSAIL) and co-author of a paper on this approach.
    Joining Schäfer on the paper are first author Julian Arnold, a graduate student at the University of Basel; Alan Edelman, applied mathematics professor in the Department of Mathematics and leader of the Julia Lab; and senior author Christoph Bruder, professor in the Department of Physics at the University of Basel. The research is published today in Physical Review Letters.
    Detecting phase transitions using AI

    While water transitioning to ice might be among the most obvious examples of a phase change, more exotic phase changes, like when a material transitions from being a normal conductor to a superconductor, are of keen interest to scientists.
    These transitions can be detected by identifying an “order parameter,” a quantity that is important and expected to change. For instance, water freezes and transitions to a solid phase (ice) when its temperature drops below 0 degrees Celsius. In this case, an appropriate order parameter could be defined in terms of the proportion of water molecules that are part of the crystalline lattice versus those that remain in a disordered state.
    In the past, researchers have relied on physics expertise to build phase diagrams manually, drawing on theoretical understanding to know which order parameters are important. Not only is this tedious for complex systems, and perhaps impossible for unknown systems with new behaviors, but it also introduces human bias into the solution.
    More recently, researchers have begun using machine learning to build discriminative classifiers that can solve this task by learning to classify a measurement statistic as coming from a particular phase of the physical system, the same way such models classify an image as a cat or dog.
    The MIT researchers demonstrated how generative models can be used to solve this classification task much more efficiently, and in a physics-informed manner.
    The Julia Programming Language, a popular language for scientific computing that is also used in MIT’s introductory linear algebra classes, offers many tools that make it invaluable for constructing such generative models, Schäfer adds.

    Generative models, like those that underlie ChatGPT and Dall-E, typically work by estimating the probability distribution of some data, which they use to generate new data points that fit the distribution (such as new cat images that are similar to existing cat images).
    However, when simulations of a physical system using tried-and-true scientific techniques are available, researchers get a model of its probability distribution for free. This distribution describes the measurement statistics of the physical system.
    A more knowledgeable model
    The MIT team’s insight is that this probability distribution also defines a generative model upon which a classifier can be constructed. They plug the generative model into standard statistical formulas to directly construct a classifier instead of learning it from samples, as was done with discriminative approaches.
    “This is a really nice way of incorporating something you know about your physical system deep inside your machine-learning scheme. It goes far beyond just performing feature engineering on your data samples or simple inductive biases,” Schäfer says.
    This generative classifier can determine what phase the system is in given some parameter, like temperature or pressure. And because the researchers directly approximate the probability distributions underlying measurements from the physical system, the classifier has system knowledge.
    This enables their method to perform better than other machine-learning techniques. And because it can work automatically without the need for extensive training, their approach significantly enhances the computational efficiency of identifying phase transitions.
    At the end of the day, similar to how one might ask ChatGPT to solve a math problem, the researchers can ask the generative classifier questions like “does this sample belong to phase I or phase II?” or “was this sample generated at high temperature or low temperature?”
    Scientists could also use this approach to solve different binary classification tasks in physical systems, possibly to detect entanglement in quantum systems (Is the state entangled or not?) or determine whether theory A or B is best suited to solve a particular problem. They could also use this approach to better understand and improve large language models like ChatGPT by identifying how certain parameters should be tuned so the chatbot gives the best outputs.
    In the future, the researchers also want to study theoretical guarantees regarding how many measurements they would need to effectively detect phase transitions and estimate the amount of computation that would require.
    This work was funded, in part, by the Swiss National Science Foundation, the MIT-Switzerland Lockheed Martin Seed Fund, and MIT International Science and Technology Initiatives. More