Graph-Based Feature Engineering to Predict the Dynamical Properties of Condensed Matter

Abstract

We present a graph theory-based method to characterise flow defects and structural shifts in condensed matter. We explore the connection between dynamical properties, particularly the recently introduced concept of ''softness'', and graph-based features such as centrality and clustering coefficients. These topological features outperform conventional features based on Euclidean metric in predicting particle mobility and allow to correctly identify phase transitions as well. These results provide a new set of computational tools to investigate the dynamical properties of condensed matter systems.