Physical interpretability of machine learning methods: learning energy barriers from local structures in supercooled fluids

Machine Learning techniques have been helpful in building our understanding of the underlying mechanisms behind local rearrangements in supercooled liquids. While some of these methods, like the 'softness' approach based on support vector machines, have successfully identified combinations of local structural attributes in a particle's environment that correlate with energy barriers to its rearrangement, the reason behind their effectiveness remains elusive. To address this, we investigate the predictive capacity and generalizability of neural networks trained to predict particle dynamics in different structural environments in the Kob-Andersen glass model. We investigate whether specific dynamical labels for rearranging particles influence the ability of various machine learning methods to generalize, and whether different machine learning methods are able to learn similar physically interpretable features, such as a connection between local structures and energy barriers.