Abstract
Advancements in computing power have made it possible to numerically simulatelarge-scale fluid-mechanical and/or particulate systems, many of which areintegral to core industrial processes. Among the different numerical methodsavailable, the discrete element method (DEM) provides one of the most accuraterepresentations of a wide range of physical systems involving granular anddiscontinuous materials. Consequently, DEM has become a widely acceptedapproach for tackling engineering problems connected to granular flows andpowder mechanics. Additionally, DEM can be integrated with grid-basedcomputational fluid dynamics (CFD) methods, enabling the simulation of chemicalprocesses taking place, e.g., in fluidized beds. However, DEM iscomputationally intensive because of the intrinsic multiscale nature ofparticulate systems, restricting simulation duration or number of particles.Towards this end, NeuralDEM presents an end-to-end approach to replace slownumerical DEM routines with fast, adaptable deep learning surrogates. NeuralDEMis capable of picturing long-term transport processes across different regimesusing macroscopic observables without any reference to microscopic modelparameters. First, NeuralDEM treats the Lagrangian discretization of DEM as anunderlying continuous field, while simultaneously modeling macroscopic behaviordirectly as additional auxiliary fields. Second, NeuralDEM introducesmulti-branch neural operators scalable to real-time modeling ofindustrially-sized scenarios - from slow and pseudo-steady to fast andtransient. Such scenarios have previously posed insurmountable challenges fordeep learning models. Notably, NeuralDEM faithfully models coupled CFD-DEMfluidized bed reactors of 160k CFD cells and 500k DEM particles fortrajectories of 28s. NeuralDEM will open many new doors to advanced engineeringand much faster process cycles.