Assessing the structural heterogeneity of supercooled liquids through community inference

Abstract

We present an information-theoretic approach inspired by distributional clustering to assess the structural heterogeneity of particulate systems. Our method identifies communities of particles that share a similar local structure by harvesting the information hidden in the spatial variation of two- or three-body static correlations. This corresponds to an unsupervised machine learning approach that infers communities solely from the particle positions and their species. We apply this method to three models of supercooled liquids and find that it detects subtle forms of local order, as demonstrated by a comparison with the statistics of Voronoi cells. Finally, we analyze the time-dependent correlation between structural communities and particle mobility and show that our method captures relevant information about glassy dynamics.