Adaptive multiscale modeling of dense granular flows using ensemble deep learning

Continuum modeling of dense granular flows in complex geometries typically relies on pre-calibrating the unknown constitutive granular rheology using discrete element method (DEM)-based rheological simulations. These computations can be expensive due to the large, a priori unknown rheological phase space that sensitively depends on particle characteristics such as friction, cohesion, and size distribution. We introduce an adaptive multiscale modeling framework that leverages ensemble deep learning to obtain on-the-fly, uncertainty-quantified approximation of the regularized constitutive rheology in a continuum simulation. Our multiscale approach adaptively samples regions of rheological phase space with high variance, thus significantly reducing the need for numerous expensive DEM simulations. The efficiency of our framework is demonstrated through continuum modeling of dense granular flows in various complex geometries, particularly investigating the role of interparticle friction and cohesion on continuum-scale granular dynamics. We evaluate the accuracy of this multiscale approach by comparison with corresponding DEM simulations. This multiscale modeling framework is extensible to a wide class of particulate materials and complex fluids beyond dense granular flows.