Assoc. Prof. Dr. Daniele Coslovich (U Trieste): Machine learning glassy dynamics: what do humans learn?

Understanding the nature of the liquid-glass transition is a long-standing open problem in condensed matter physics. Several theories have been developed to describe glassy dynamics, but none of them provides yet a fully satisfactory description. Recently, machine learning models have been shown to successfully predict the dynamical properties of glass-forming liquids from structural data [1]. Accurate predictions, however, do not guarantee an understanding of the underlying physical phenomena and the key factors that control them. In this talk, I will provide examples, drawn from both supervised and unsupervised learning, illustrating merits and limitations of data-driven models for the structure and dynamics of glass-forming liquids. In particular, I will present results for interpretability metrics of linear models of glassy dynamics built on high-dimensional structural descriptors [2], along with an extension to a broader range of datasets. The possibility of a trade-off between prediction accuracy and interpretability will be critically discussed.
[1] G. Jung et al., Nature Reviews Physics 7, 91 (2025)
[2] A. Sharma, C. Liu, M. Ozawa, D. Coslovich, Phys. Rev. Mater. 10, 035602 (2026)
Cookies, tea