BARS and mini-BARS for Iteratively Learned Sampling and Tiling of Phase Space in High Energy Density Plasma Kinetic Simulations

We will show the efficacy of Bidirectional Adaptive Refinement Scheme, BARS and mini-BARS (a reduced functionality version), to tackle plasma kinetic simulations with much more control on speed, accuracy and interpretability than traditional PIC code (Monte Carlo) sampling. We will take the initial value probelm of nonlinear kinetic electron plasma waves as an example, and show how sparse sampling in the bulk and dense sampling in the tail of the electron velocity distribution function will guarantee both sufficient accuracy and great speed at the same time. We will indicate how these methdos can be turned into full Machine Learning algorithms adapting what is learned at one set of plasma parameters to efficiently run nearby parameter cases, such as changed perturbation amplitude and wavenumber. Strict error control and novel Python diagnostics with reconstructed velocity distribution functions will be used. All our proof of principle results will be shown with DPIC, a Los Alamos Electrostatic PIC code.