When water freezes, it changes from a liquid phase to a solid phase, resulting in significant changes in properties such as density and volume. Although phase transitions in water are so common that most people have probably never thought about them, phase transitions in new materials and complex physical systems are an important area of research.
To fully understand these systems, scientists must be able to recognize the stages and detect the transitions between them. However, how to quantify phase changes in unknown systems is often unclear, especially when data are lacking.
Researchers at MIT and the University of Basel in Switzerland have applied generative artificial intelligence models to this problem, developing a new machine learning framework that can automatically map the phase diagrams of new physical systems.
Their physics-based machine learning approach is more efficient than laborious manual methods that rely on theoretical expertise. Importantly, their approach leverages generative models and therefore does not require large labeled training datasets used in other machine learning techniques.
Such a framework could help scientists explore the thermodynamic properties of new materials or detect entanglement in quantum systems. Ultimately, this technology could allow scientists to autonomously discover unknown phases of matter.
“If you have a new system with completely unknown properties, how do you choose which observable quantities to study? At least data-driven tools allow you to scan large new systems in an automated way. The hope is that it will be able to point out important changes in a system. This could become a tool in a pipeline to automatically discover new exotic properties in phases. ,” said Frank Schaefer, a postdoctoral fellow in the Julia Lab at the Computer Science and Artificial Intelligence Laboratory (CSAIL) and co-author of the following paper. this approach.
Schaefer was joined on the paper by Julian Arnold, a graduate student at the University of Basel and first author. Alan Edelman, Professor of Applied Mathematics in the Department of Mathematics and Julia Lab Leader. and lead author Christoph Bruder, professor at the Department of Physics at the University of Basel. This research today Physical review letter.
Phase transition detection using AI
While the transition of water to ice may be one of the most obvious examples of a phase change, more exotic phase changes, such as when a material transitions from a normal conductor to a superconductor, are of intense interest to scientists. are collecting.
These transitions can be detected by identifying “ordinal parameters”, which are quantities that are important and expected to change. For example, when the temperature of water drops below 0 degrees Celsius, water freezes and changes into a solid state (ice). In this case, the appropriate order parameter can be defined in terms of the fraction of water molecules that are part of the crystal lattice and the fraction that remains in a disordered state.
Previously, researchers relied on physics expertise to manually construct phase diagrams with a theoretical understanding of which order parameters are important. Not only is this tedious for complex systems and probably impossible for unknown systems with new behavior, but it also introduces human bias into the solution.
More recently, researchers have used machine learning to show that such models can classify measurement statistics as coming from a particular stage of a physical system, in the same way that they classify images as cats or dogs. By learning, we are beginning to build discriminative classifiers that can solve this task.
MIT researchers have demonstrated how generative models can be used to solve this classification task more efficiently and in a physics-based way.
The Julia programming language, a popular language for scientific computing that is also used in MIT's introductory linear algebra classes, offers many invaluable tools for building such generative models, Schaefer said. adds.
Generative models, such as those underlying ChatGPT and Dall-E, typically work by estimating a probability distribution for some data, and then selecting new data points that fit that distribution (new cat images similar to existing cat images). image). .
However, when simulations of physical systems using proven scientific methods are available, researchers can obtain models of their probability distributions for free. This distribution represents the measurement statistics of a physical system.
More knowledgeable model
According to the MIT team's insight, this probability distribution also defines a generative model for building classifiers. Rather than learning a classifier from samples, as was done in discriminative approaches, they incorporate generative models into standard statistical formulas to directly construct classifiers.
“This is a really great way to incorporate what we know about physical systems deep into machine learning schemes. It's more than just doing feature engineering on data samples and simple inductive biases.” Schaefer says.
This generative classifier can determine what stage a system is in given parameters such as temperature and pressure. The classifier also has system knowledge because the researcher is directly approximating the probability distribution underlying the measurements from the physical system.
This allows their method to perform better than other machine learning methods. Their approach also greatly improves the computational efficiency of identifying phase transitions, as it works automatically without the need for extensive training.
After all, just like asking ChatGPT to solve a math problem, researchers can ask generative classifiers questions like “Does this sample belong to Phase I? Does it belong to Phase II?” You can to be. or “Was this sample produced at a high temperature or a low temperature?”
Scientists use this approach to solve a variety of binary classification tasks in physical systems, to detect entanglement (states are entangled or not) in quantum systems, and to solve problems such as whether theory A or B You can also decide which is the best solution for your problem. This approach can also be used to make large language models like ChatGPT more useful by identifying how certain parameters should be tuned for the chatbot to provide the best output. It can be deeply understood and improved.
In the future, the researchers also hope to study theoretical guarantees for estimating the number of measurements and amount of computation required to effectively detect phase transitions.
This research was funded in part by the Swiss National Science Foundation, the MIT-Swiss Lockheed Martin Seed Fund, and the MIT International Science and Technology Initiative.
