Use of Hybrid Architectures in Particle-in-cell Simulations using GPUs*

A recent report by the National Academy cites the importance of kinetic, particle-in-cell approaches that use Graphical Process Unit (GPU) computing to take advantage of large processor core counts for modeling etching and microelectronics fabrication in plasma reactors. However, they also acknowledge the challenge of managing thousands of separate threads of computation. This talk will discuss our approach which centers around the use of our three-dimensional GPU Cuda-based Hybrid Approach for Octree Simulation (CHAOS) Particle-in-Cell (PIC)/ direct simulation Monte Carlo (DSMC) simulation tool. Our general concept is to use a hierarchical octree structure to group nearest neighboring particles into leaf nodes, which are analogous to cells in structured grid-based solvers. Separate octree grids are then used for the DSMC and PIC portions of the calculation since the DSMC and PIC use different refinement criteria based on the local mean free path and Debye length, respectively. In many of the plasmas that we have modeled, the Debye length is much smaller than the local mean free path so that if the former were used as the refinement criteria the DSMC leaf nodes would be over refined leading to the prohibitive increase in the required number of macroparticles and therefore, computational cost. We have verified the implementation and accuracy of our PIC module in CHAOS by comparing results of a mesothermal collisionless plasma plume with other published results and since the original software development began, the capabilities of CHAOS have grown to include the modeling of ion thruster backflow contamination and sputtering, sputtering in ground vacuum chambers, secondary electron emission (SEE), the study of the role of electrostatic solitary waves in ion beam neutralization, and magnetic nozzles. The paper will discuss the upcoming challenges as we push kinetic methods to more complex systems as well as the future opportunities.