Accelerating Ensembles of Plasma Kinetic Simulations through Adaptive Learning: NSCAR, BARS, mini-BARS, TPOR and all that

We will describe elements of NSCAR: Nearby Skeleton Constrained Accelerated Recomputing and its proginy BARS, mini-BARS and TPOR. Nearby solutions are compressed and added as a variational constraint on the overall Lagrangian formulation of the Kinetic modeling problem. The constraints force the search for solutions to favor proximity to the previously obtained solutions. BARS stands for Bidirectional Adaptive Reduction Scheme. This adaptively resampled phase-space algorithm that traverse forwards and backwards in time enables new variational optimization that benefits ensembles of simulations. The key is to use similar patches in phase space to provide excellent initial guesses for optimum computational performance in nearby parameter space. This way, what is learned with O(4x) overhead in one simulation is shared between nearby runs and becomes essentially free. A simplified version of BARS, called mini-BARS, has been used together with the PIC code DPIC to study ponderomotively driven electron plasma waves and KEEN waves with velocity-partitioned phase space (with variable-width velocity strips not general patches in phase space) and found to accelerate computations by a factor of O(100), at very low overhead cost (roughly 4x).

The winning strategy is to learn how to over-sample the tail and undersample the bulk of the velocity distribution function adaptively.