Machine-learned structure/dynamics relation in sheared jammed packings

In disordered systems, using local structure to identify which particles are likely to rearrange under thermal fluctuations or applied load has been a longstanding challenge. Recently, machine learning has been used to construct a local structural variable, "softness", that is highly predictive of rearrangements in several disordered systems. Here we describe modifications made in the analysis that simplify interpretation and raise training accuracy for athermal packings of soft spheres under quasistatic shear. We obtain a "softness" that is highly predictive of the rearrangements at the onset of instabilities. Furthermore, we show that for jammed Hertzian packings, softness can be represented simply in terms of gaps and contacts between neighboring particles. We show how this picture depends on pressure above jamming and spatial dimension.