Andreas Mayr, Sebastian Lehner, Arno Mayrhofer, Christoph Kloss, Sepp Hochreiter, and Johannes Brandstetter
The paper presents a new method based on graph neural networks (GNNs) for modeling 3D granular material flow with complex boundary conditions. This work is important for understanding particle flow dynamics and for designing industrial processes. As there are no underlying governing equations for granular material flow, the authors employ simulations. The new method is tested on simulations of physical systems found in real world applications, such as particle dynamics in a rotating drum and particle flow through a hopper. The simulation data is generated with the open-source software LIGGGHTS based on the discrete element method (DEM). Time-transition model is employed, consisting of an encoder, a processor, and a decoder to predict particle accelerations. The authors discuss implementing GNNs to model 3D objects, boundary conditions, particle-particle, and particle-boundary interactions. Triangularization of 3D objects is achieved by inserting virtual nodes if particles are close to boundaries and computing proximity to the closest triangle. The network predictions are invariant with respect to the orientation of the normal vectors representing the planes of triangle walls. Particle trajectories obtained with the trained GNN are compared to the simulation trajectories obtained by LIGGGHTS. The results show the expected increase of the mixing entropy over time and a good agreement for aggregated (time- and particle-averaged) properties.
arXiv: 2105.01636, 2021-05-04