425229 Using GPUs to Simulate Three-Dimensional Segregating Granular Flows

Monday, November 9, 2015: 8:30 AM
254B (Salt Palace Convention Center)
Austin B. Isner1, Paul B. Umbanhowar2, Julio M. Ottino1,3 and Richard M. Lueptow2,3, (1)Chemical and Biological Engineering, Northwestern University, Evanston, IL, (2)Mechanical Engineering, Northwestern University, Evanston, IL, (3)The Northwestern Institute on Complex Systems (NICO), Northwestern University, Evanston, IL

Segregation driven by differences in particle size or density in flowing granular mixtures presents numerous challenges in the industrial processing of bulk solids. Recent continuum-based models, informed by measurements from discrete element method (DEM) simulations, are promising in their ability to quantitatively predict particle species concentration as a function of both space and time, given the flow velocity field and kinematic parameters such as the diffusion coefficient. However, these models have thus far been applied only to 2D or “quasi”-2D flows characterized by a small third dimension of O(10) particle diameters, such as bounded heaps, chute flows, or rotating tumbler flow. Fully three-dimensional flows constitute an even broader class of granular segregation problems; however, they are complicated by experimental difficulties in accessing the flow kinematics in the interior of granular mixtures. Numerical methods such as DEM can help by offering detailed information required for model validation, but are almost always burdened by the computational complexity posed by even small 3D systems. Here, we discuss the role of graphics processing units (GPUs) as a powerful entry-level alternative to address the high computational costs associated with simulating fully 3D granular systems with relatively modest particle counts (O(106-107) particles). The GPU-based simulation framework is applied to two cases of 3D segregating granular flows. In the first case, we employ a DEM to simulate the filling of a 3D cylindrical silo with a size bidisperse granular mixture. The second case considers a three-dimensional sheared assembly of highly size-polydisperse spheres. In each case, the results of particle segregation are compared with predictions from the continuum theory. Algorithmic optimizations for contact detection and nearest neighbor search are identified which are crucial to the efficient implementation of many-particle dynamics on the GPU. Funded by The Dow Chemical Company.

Extended Abstract: File Not Uploaded