Robust Prediction of Force Chains in Jammed Solids using Graph Neural Networks

TL;DR

This paper demonstrates that graph neural networks can accurately predict the formation of force chains in jammed solids from local structural information, even without force data, and are robust across various system parameters.

Contribution

The study introduces a GNN-based method to predict force chains in jammed materials solely from structural data, showing robustness and scalability.

Findings

GNN accurately predicts force chains without force measurements.

Prediction robustness across different packing fractions and interactions.

Method scalable to larger systems than trained on.

Abstract

Force chains, which are quasi-linear self-organised structures carrying large stresses, are ubiquitous in jammed amorphous materials, such as granular materials, foams, emulsions or even assemblies of cells. Predicting where they will form upon mechanical deformation is crucial in order to describe the physical properties of such materials, but remains an open question. Here we demonstrate that graph neural networks (GNN) can accurately infer the location of these force chains in frictionless materials from the local structure prior to deformation, without receiving any information about the inter-particle forces. Once trained on a prototypical system, the GNN prediction accuracy proves to be robust to changes in packing fraction, mixture composition, amount of deformation, and the form of the interaction potential. The GNN is also scalable, as it can make predictions for systems much larger than those it was trained on. Our results and methodology will be of interest for experimental realizations of granular matter and jammed disordered systems, e.g. in cases where direct visualisation of force chains is not possible or contact forces cannot be measured.