A Particle Segregation Study Using GPU-Accelerated DEM Simulations with Elastic Tangential Friction

The discrete element method (DEM) is a popular simulation technique that has proven useful in the study of granular flows, particularly their assembly microstructure, interparticle force networks, stress distribution, and flow kinematics. Of practical concern is the simulation of particle systems at industrially relevant length and time scales, which may require simulating tens to hundreds of millions of particles using a variety of flow geometries and contact force models. Although progress has allowed these computationally expensive simulations to run on highly scalable multi-processor systems, this approach may require significant capital investment or access to high performance computing resources. However, general purpose graphics processing unit (GPU) computing approaches have recently become a viable alternative that leverages the built-in parallelism of discrete GPU hardware, which can contain many thousands of parallel processing cores on a single device, as opposed to the four or six cores found in many modern CPU’s. Here, we present a proof-of-concept implementation of DEM on a consumer-grade GPU based on Nvidia’s CUDA Particles SDK, which incorporates a linear spring dashpot normal force model and a sliding and elastic tangential frictional force model. The DEM simulations are validated by comparing particle streamwise concentration profiles in quasi-2D bounded heap flows of bidisperse mixtures to those of experiments. We investigated the dependence of several kinematic parameters including flowing layer depth, diffusivity, and percolation length on various control parameters such as particle size ratio and two-dimensional feed flow rate. In particular, we find that the interparticle percolation length scale normalized by the small particle diameter is approximately proportional to the logarithm of particle size ratio, R, (R = dl / ds, where dl and ds are the diameters of the large and small particle, respectively) for size ratios in the range R = 1.1 to R = 3. An approach to generalize our understanding of shear-driven segregation to polydisperse systems is also tested using GPU-accelerated DEM simulations of polydisperse mixtures. As well as achieving real-time visualization of simulation results, we achieve a performance improvement that is on average one order-of-magnitude greater than an analogous DEM algorithm run on a traditional multi-core processor. The cost-savings (on a TeraFLOPS/$ basis) of GPU-accelerated DEM makes GPU-based DEM a promising platform for sustained scalability in simulations of certain granular systems.