When water freezes, it changes from a liquid phase to a solid phase, resulting in drastic changes in properties such as density and volume. Phase transitions in water are so common that most of us probably don't even think about them, but phase transitions in new materials or complex physical systems are an important area of study.
To fully understand these systems, scientists must be able to recognize the phases and detect the transitions between them. But how to quantify phase changes in an unknown system is often unclear, especially when data are scarce.
Researchers at MIT and the University of Basel in Switzerland applied generative models of artificial intelligence to this problem, developing a new machine learning framework that can automatically map phase diagrams for new physical systems.
Their physics-informed machine learning approach is more efficient than laborious, manual techniques that rely on theoretical expertise. Importantly, because their approach leverages generative models, it does not require the large, labeled training datasets used in other machine learning techniques.
Such a framework could help scientists investigate the thermodynamic properties of novel materials or, for example, detect entanglement in quantum systems. Ultimately, this technique could make it possible for scientists to autonomously explore unknown phases of matter.
“If you have a new system with completely unknown properties, how do you choose which observables to study? The hope, at least with data-driven tools, is that You can automatically scan large new systems, and it will alert you to important changes in the system, says a postdoc in the Julia Lab at the Computer Science and Artificial Intelligence Laboratory (CSAIL) and a paper co-author. This approach could be a tool in the pipeline for automated scientific discovery of new, exotic properties, says Frank Schaeffer.
The first author to join Schäfer on the paper is Julian Arnold, a graduate student at the University of Basel. Alan Edelman, professor of mathematics in the Department of Mathematics and leader of the Julia Lab; and senior author Christoph Broder, professor at the University of Basel's Department of Physics. The research is published today. Physical examination letters.
Phase transition detection using AI
While the water-to-ice transition may be one of the most obvious examples of a phase transition, more exotic phase transitions, such as when a material transitions from an ordinary conductor to a superconductor, are of intense interest to scientists. happens.
These transitions can be detected by identifying the “order parameter”, a quantity that is important and expected to change. For example, water freezes and transitions into a solid phase (ice) when its temperature drops below 0 degrees Celsius. In this case, an appropriate order parameter can be defined in terms of the ratio of water molecules that are part of the crystal lattice versus those that remain in the undissociated state.
In the past, researchers have relied on physics expertise to manually draw phase diagrams to figure out which order parameters are important. Not only is this inconvenient for complex systems, and perhaps impossible for unknown systems with new behaviors, it also introduces human bias into the solution.
More recently, researchers have begun to use machine learning to create discriminative classifiers that can solve this task by learning to classify measurement statistics coming from a particular phase of a physical system, such as Such models classify an image as either a cat or a dog.
MIT researchers demonstrated how generative models can be used to solve this classification task 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 building such creative models, Shaffer added.
Generative models, such as those underlying 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. are (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 to sample its probability distribution for free. This distribution describes the measurement statistics of physical systems.
A more knowledgeable model
The MIT team's insight is that this probability distribution also defines a generative model on which classification can be made. They plug generative models into standard statistical formulas to make classifications directly rather than learning from samples, as was done with discriminant approaches.
“It's a really good way to incorporate what you know about your physical system into your machine learning scheme. It goes far beyond performing feature engineering on your data samples or simple attractive biases, ” says Schäfer.
This generative classification can determine which phase a system is in given some parameter such as temperature or pressure. And since researchers directly estimate the probability distribution of the underlying measurements from the physical system, the classifier has knowledge of the system.
This enables their method to outperform other machine learning techniques. And because it can operate automatically without the need for extensive training, their approach significantly increases the computational efficiency of phase transition identification.
At the end of the day, just as one might ask ChatGPT to solve a math problem, researchers can ask the generative classifier questions like “Is this sample from Phase I or Phase II?” or “Was this sample prepared at a high temperature or a low temperature?”
Scientists can also use this approach to solve various binary classification tasks in physical systems, possibly to detect entanglement in quantum systems (is the state entangled or not?) or it can determine whether theory A or B is best suited to solving a particular problem. They can also use this approach to better understand and improve large language models like ChatGPT by identifying how certain parameters should be tuned so that the chatbot gives the best results.
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 be required.
This work was funded by the Swiss National Science Foundation, the MIT-Switzerland Lockheed Martin Seed Fund, and the MIT International Science and Technology Initiatives.