Graph neural network prediction of displacements in sheared granular systems

Jammed granular systems are disordered solids with resistance to compression and shear.  Recently, machine learning (ML) has successfully predicted jammed softness and force network structure. Here, we ask if molecular dynamics (MD) simulation may be replaced by ML in order to track motion. We study sheared systems with or without volume-excluding pins, which can fine-tune structure.

 

We predict static distances and displacements in a sheared 2d system of soft, bidisperse particles with a graph neural network (GNN). Nodes contain particle position and size information; edges are drawn within a threshold interparticle distance; and the MD data are obtained using LAMMPS.

 

Relative distances to walls and pins are suitable tasks for tuning model hyperparameters. Predicting displacements, we explore the time attenuation of accuracy as the system evolves from the initial state. Future work might involve adding particle velocities to the feature vector, thus extending prediction to arbitrary MD timesteps.