Data-driven model order reduction for granular media

Abstract We investigate the use of reduced-order modelling to run discrete element simulations at higher speeds. Taking a data-driven approach, we many offline in advance and train model predict velocity field from mass distribution system control signals. Rapid inference particle velocities replaces intense process computing contact forces updates. In coupled DEM multibody simulation, predictor can be trained output interfacial reaction as well. An adaptive order reduction technique is investigated, decomposing media domains solid, liquid, gaseous state. The applied solid liquid where motion strongly correlated with mean flow, while resolved used for domains. Using ridge regression predictor, performance tested on pile discharge bulldozing. measured accuracy about 90% 65%, respectively, speed-up range between 10 60.