Physics informed neural network model for multi-particle interaction forces

Abstract

The discrete element method (DEM) model calculates interaction forces between each pair of particles. However, it becomes computational expensive especially when the number of particles is large. In this study, a novel artificial neural network (ANN) model is proposed to replace the model of interaction forces between multiple particles in DEM including contact force and electrostatic force. The ANN model combines the residual network (ResNet) with the physics informed neural network (PINN). The physical loss term is derived from the Newton's third law about internal forces in multi-particle system. The performance of the ANN model is evaluated based on the DEM simulation data of 100, 200, and 300-particle system in a wall-bounded 2D swirling flow. It is found that the computing time is reduced nearly an order of magnitude (7–10 times) compared with the DEM model. In addition, the accuracy of the ANN model achieves the $R^2>0.93$ with only $\le 2\%$ particles are not well predicted.