Particle trajectory prediction in discrete element simulations using a graph-based interaction-aware model

This study explores the applicability of a graph-based interaction-aware trajectory prediction model, originally developed for the transportation domain, to forecast particle trajectories in three-dimensional discrete element simulations. The model and our enhancements are validated at two typical particle simulation use cases: (i) particle flow in a representative unit cell with periodic boundary conditions (PBCs) in combination with sinusoidal velocity profile and (ii) shear flow in a representative unit cell with Lees–Edwards boundary conditions (LEBCs). For the models to learn the particle behavior subjected to these boundary conditions requires additional data transformation and feature engineering, which we introduce. Furthermore, we introduce and compare two novel training procedures for the adapted prediction model, which we call position-centric training (PCT) and velocity-centric training (VCT). The results show that the models developed for the transportation domain can be adapted to learn the behavior of particles in discrete element simulations. • Extended the usage of the trajectory prediction model from the transportation domain. • Predicted particle trajectories in three-dimensional discrete element simulation. • Introduced feature engineering for periodic and Lees–Edwards boundary conditions. • Enhanced accuracy by introducing position and velocity-centric training strategies.