MAPping dynamic heterogeneity in supercooled glass-formers

A central question in glass physics is whether dynamic heterogeneity of supercooled liquids can be inferred from static structure. Successful models based on supervised and unsupervised machine-learning predict mobility from particle positions but either require labels in training and/or do not offer simple interpretation. Here, we propose to diagnose dynamic heterogeneity using an autoregressive generative model called MAP that learns probabilities of particle configurations conditioned on fragmented local environments from molecular-dynamics data. We verify that in its generative capacity, MAP reproduces key benchmarks for the supercooled Kob-Andersen binary system, such as the relevant pair-correlation functions and the point-to-set distance temperature scaling. We then explore the behavior of the MAP-derived reaction coordinate Ω and find that its modulations coincide with dynamic excitations. We tune Ω into a hypersensitive regime and find that it reports more broadly on local dynamic events, which we characterize as concurrent nearby excitations. This unsupervised-learning signal links rare events to atypical states of fragmented local environments, and it offers a new intuitive way to characterize and explore the dynamic heterogeneity in supercooled glass-formers.