Research On A Simulation Method For Spherical Particles In A Ball Mill Based On DEM And GPU

The movement of the granular system is an important process in industrial production.It is widely used in industries such as mining,pharmaceutical industry and electric power,which are related to the national economy and the people’s livelihood.However,the driving energy consumption is huge and the utilization rate of the driving energy is relatively low.Industrial ball mill is a typical example of its industrial applications.Studying the motion characteristics of the ball mill media and materials and understanding the mechanism of various phenomena in the mill has a very important effect on improving the energy utilization rate of the ball mill.Computer simulation based on DEM is the main way to achieve this goal,however,due to the large scale of the medium in the ball mill and the extremely high computational complexity of the DEM,large-scale simulations requires much more computing resources and a lot of computing time.It is difficult to make the simulation of the particle system a daily means for theoretical researching.Therefore,this paper combines DEM and GPU parallel computing architecture to realize a GPU-based partical motion simulation in a ball mill,and hopes to improve the speed of computer simulation with more common equipment,and provide a faster way for this research area.Firstly,a theoretical model based on DEM for the motion simulation of ball mill media is established.By using the CUDA parallel computing model,the algorithm implementation of the DEM-based ball mill particle motion simulation model on the GPU is established.And each GPU thread accounts for the computing task of each particle.By decomposing the various calculation steps of the DEM,the calculation task of each calculation step is simplified.Secondly,The space binning of the particles is decomposed into three steps including hashing,sorting and adjacent judgment and boxing.The simpler calculation method is used to realize the potential neighbor search of the particles,so that the related calculation tasks of each particle are greatly reduced,which is more suitable for the lightweight thread structure of the GPU;For the uniform size particle system,by using the mask search method,half of the search area of the particles are reduced and the repeated search is removed,further reducing the calculation time of the contact judgment process and improving the calculation efficiency;For particle systems with uneven size distribution,the bounding box method for spatial decomposition and binning is used,which can dynamically adjusts the neighboring regions of the particles according to target particle size and position.Moreover,a large number of unnecessary contact judgment caused by the large increase in the number of particles per cell unit is reduced,further improving the computational efficiency of the coarse phase contact search.Thirdly,Due to the fact that it is difficult to apply the tangential contact force model based on contact history to the GPU architecture,which will greatly reduce the computational efficiency,the simple friction model that ignores the tangential elastic deformation process is selected which can reduce the complexity of contact force calculation and saves more device memory;By simplifying the geometry,the contact between the particles and the geometry is simplified,thus avoiding a large number of additional discrete units caused by the discretization of the aggregate.By judging the relative position of the particles and the geometry,a large number of particles that are not in real contact with the geometry are eliminated,and the contact judgment between the particles and the geometry is reduced.Finally,by comparing the simulation results of the commercial software using the CPU parallel computing method and the experimental results of the small roller experiment,the accuracy of the calculation model proposed in this paper is verified.Then,by adopting three simulation schemes,under the hardware conditions in this paper,the computational efficiency of the GPU parallel computing method is compared with that of the CPU parallel computing method.By comparing and analyzing the computing time under different thread structures and different spatial grid sizes,the optimal thread structure range and grid size selection range is found.