GRAPH NEURAL NETWORKS ACCELERATED GRANULAR FLOW BASED ON DISCRETE ELEMENT METHOD

This study aims to explore a new model based on graph neural network (GNN) to accelerate discrete element method (DEM) calculations, so as to improve the computational efficiency and accuracy of granular flow simulation. Although the traditional DEM method is accurate, it is computationally expensive. GNNs have a natural advantage in simulating DEM. In the GNNs, particles are represented as nodes and the interactions between particles are represented as edges. The proposed acceleration model combined two types of GNNs, namely particle-particle graph neural network (P-P GNN) and particle-boundary graph neural network (P-W GNN), which can learn the information of particle-particle and particle-boundary contact relationships, respectively. The effectiveness and superiority of the model were verified by simulating three granular flow scenarios: horizontal roller, inclined roller, and hopper simulation. The results show that the GNN model effectively captured the complex contact relationship in granular flow, greatly improved the calculation speed, and achieved about 30 times acceleration compared with traditional DEM. Secondly, the model show high accuracy in one-step prediction under different particle flow scenarios, and performed well in the prediction of macroscopic features, and can accurately predict the angle of repose, temperature, and center of gravity changes of particles. In addition, this paper also studied the influence of hyperparameters on the prediction results, such as cutoff distance and number of GNN layers. The appropriate cutoff distance can limit particles from crossing the boundary. The number of GNN layers between 3 and 10 has no significant effect on the prediction results. This study provided a research basis for further optimizing the GNN model for granular flow simulation. Further research should be conducted on the impact of different hyperparameters on GNN prediction results, considering the combination effect of multiple hyperparameters to obtain the optimal setting. In addition, how to make GNN learn the influence of factors such as particle size and shape is also worth considering. © 2024 Chinese Society of Theoretical and Applied Mechanics. All rights reserved.