Scientists use generative AI to answer complex questions in physics


Scientists from MIT and the University of Basel in Switzerland have introduced a novel machine-learning framework that employs generative artificial intelligence (AI) models to automatically map out phase diagrams for novel physical systems. This groundbreaking approach addresses the challenge of quantifying phase changes in complex systems with limited data.

Phase transitions, such as the freezing of water, are commonplace, but detecting phase changes in novel materials or intricate physical systems presents unique challenges. Traditional manual techniques rely heavily on theoretical expertise and can be time-consuming. To overcome these limitations, the researchers turned to generative AI models to develop a more efficient and data-driven approach.

Their framework, detailed in a paper published in Physical Review Letters, leverages generative models to recognize phases and detect transitions in physical systems. Unlike conventional machine-learning techniques that require extensive labeled datasets, this approach utilizes physics-informed machine learning and does not depend on large training datasets.

The researchers demonstrated the effectiveness of their method in detecting phase transitions by identifying order parameters that signify changes in the system. By incorporating knowledge about the physical system directly into the machine-learning scheme, the framework outperforms traditional techniques and enhances computational efficiency.

Moreover, this approach opens up possibilities for various binary classification tasks in physical systems, such as detecting entanglement in quantum systems or selecting the most suitable theoretical model for a given problem. It could also contribute to improving large language models like ChatGPT by optimizing parameters for better performance.

Looking ahead, the researchers aim to explore theoretical guarantees regarding the number of measurements required to detect phase transitions effectively and estimate the computational resources needed for implementation.

Funding for this research was provided by the Swiss National Science Foundation, the MIT-Switzerland Lockheed Martin Seed Fund, and MIT International Science and Technology Initiatives.