Researchers from MIT and the University of Basel in Switzerland have developed a new machine-learning framework that can map phase diagrams for novel physical systems automatically. By applying generative artificial intelligence models, the team has developed a more efficient method for tracking and understanding phase transitions in water and other complex physical systems, which offers significant advancements in fields such as the study of thermodynamic properties of unknown materials or detecting entanglement in quantum systems.
Phase transitions refer to the change of a system from one state of matter to another, such as liquid to solid. Historically, these transitions have been manually detected and mapped using human expertise and theoretical concepts. However, this traditional method is laborious and can also introduce human bias into the solution. In recent years, machine learning has been used to build classifiers that can sort data into different categories automatically, offering a more efficient alternative.
The work of the team from MIT and the University of Basel takes these machine learning developments a step further by using generative models. Unlike other classifier models, generative models can produce new data points that fit within its learned probability distribution. The team realized this distribution actually represents a generative model upon which a classifier can be constructed. This classifier can then determine what phase the system is in given certain inputs like temperature or pressure, doing so much more quickly and accurately than other techniques because it directly leverages the probability distributions underlying measurements from the physical system.
This technique provides the scientific community with a powerful new tool to explore and understand complex physical systems and their behaviours. By asking the generative classifier specific questions, researchers can gain greater insights into the behavior of certain systems or materials under different conditions. This could be particularly beneficial for quantum physics or for assessing which of multiple theories is most effective for addressing a given problem.
The team also hopes to gain further insights into how to optimize the performance of large AI models using this technique in the future. For example, they could determine how certain parameters in a language-processing algorithm like ChatGPT should be changed to improve its outputs. However, more theoretical work is still required to determine how many measurements are needed to accurately detect phase transitions and how much computational power such a process would require.
Their research, funded by the Swiss National Science Foundation, the MIT-Switzerland Lockheed Martin Seed Fund, and MIT International Science and Technology Initiatives, has been published in Physical Review Letters.
