The prediction of dynamical quantities in granular avalanches based on graph neural networks.

This article focuses on using graph neural networks (GNNs) to predict dynamic features of three-dimensional (3D) granular avalanches in rotating drums based solely on their initial static microstructures. By training GNN models on large-scale numerical simulations performed with the LAMMPS package, the study accurately predicts grain-scale velocities and nonaffine displacements—an indicator of plastic rearrangements—with Pearson correlation coefficients around 0.8. The models demonstrate robustness across variations in interaction potentials, particle size polydispersity, and coordinate noise, and reveal a strong correlation between global velocity and velocity fluctuations in 3D granular avalanches. These findings suggest that GNNs can effectively capture complex avalanche dynamics in granular materials, offering potential applications in understanding natural hazards and industrial processes involving granular flows.