Estimating the entropy of non-equilibrium physical, chemical, and biological systems from snapshot images using ML techniques


I am interested in interdisciplinary studies between statistical physics and machine learning. Nowadays, machine learning is becoming a powerful tool for physicists to analyze various physics problems. Conversely, the concepts and techniques developed in physics help us uncover the mystery behind the outstanding performance of machine learning architectures, which have been considered as black boxes. Currently, we are developing machine learning methods inspired by statistical physics, making synergy between physics and machine learning.

Generative models are making a huge impact on society, such as image generation and chat GPT, to name but a few. Although they have been showing remarkable successes, the interpretability of the outstanding performance is poorly understood. Instead, they are rather considered as a black-box. We are working on interdisciplinary studies between statistical physics and generative models in order to understand the mechanism of learning and to construct efficient generation algorithms. In particular, he is developing generative models based on a renormalization group technique. Renormalization group is a powerful technique in modern theoretical physics, which allows us to process data hierarchically in a scale-by-scale manner. This new generative model allows us to interpret the learning process as well as results in a physically intuitive way, and provides us with a new fast Monte-Carlo sampling algorithm to generate new examples.


Entropy plays an important role in science and engineering as it characterizes various features of systems, such as the time evolution of a macroscopic sample, phase transformation, the emergence of certain orders, etc. However, measuring entropy is a non-trivial task, in particular when the system falls out of equilibrium. This stems mainly from the difficulty of obtaining a high-dimensional probability distribution from data (e.g., experimental images, simulated configurations). We are developing a scheme to measure Shannon entropy by using the generative model mentioned above. The method will be applied to image data from various domains, including physics, chemistry, and biology.

Published on  January 25, 2024
Updated on January 25, 2024