NeuralDEM - Real-time Simulation of Industrial Particulate Flows
In this work, colleagues from NXAI (now Emmi AI) and I let universal physics transformers learn the dynamics of granular matter both in the dense, pseudo-steady regime of hopper flow and the dilute, dynamic behavior encountered in fluidized bed reactors. The neural networks were trained with a large number of DEM and CFD-DEM trajectories for varying particle properties, boundary conditions and geometries. Long-term predictions for different values of these properties showed (i) simulation speed ups of several orders of magnitude, (ii) very good generalization capabilities to unseen conditions, and (iii) the feasibility to directly use macroscopic material properties (e.g. angle of repose) instead of microscopic DEM parameters.
Nov 14, 2024
