Simulation of heterogeneous discrete element particle motion based on graph neural networks

Deep learning has shown great potential in simulating the physical dynamics of heterogeneous discrete elements. Recent works have further enhanced performance by combining physical laws with deep learning. However, challenges arise when the discrete element particles have inconsistent sizes, irregular large particles, or slow particle motion. These challenges include reduced simulation accuracy, increased computational complexity, difficulties in model construction, increased data requirements and decreased simulation efficiency. The existing methods require more complex models, more data for training, and more computational resources and time. To address these issues, we propose a Heterogeneous Discrete Element Graph Network (HDE-GN) designed to learn the dynamics of heterogeneous discrete elements more accurately and efficiently. We incorporate physical prior knowledge into the model, designing vertex and edge embedding processes to transform heterogeneous discrete element data into graph information. The model’s information transmission framework is designed to update vertex and edge feature vectors through information construction, aggregation and vertex update processes, thereby improving the model’s accuracy and enabling more complex and realistic physical simulations. Experimental evaluations of datasets demonstrate that our model achieves state-of-the-art prediction accuracy, effectively simulating the motion of heterogeneous discrete element particles.