Grad Talk: Using machine learning to accelerate particle in cell simulations

Abstract: In many high energy astrophysical sources, like Active Galactic Nuclei and Supernova Remnants, magnetohydrodynamic flows play a central part in the acceleration of particles and production of electromagnetic radiation. While observations typically only reveal the large-scale features of these flows, their physical origin might emerge at the kinetic scale, the scale of the motion of individual particles. Therefore, theoretical modeling of such systems requires covering several orders of magnitude. The standard way to bridge this gap is through particle in cell (PIC) simulations.

PIC simulations are, however, computationally expensive. Typical modern 2D PIC codes are able to simulate boxes of length scales 8 - 9 orders of magnitude smaller than the actual physical systems, and over timescales much smaller than the hydrodynamic timescale. It is thus evident that large leaps in the numerical efficiency of our methods are needed, especially given the increasing quality of observational data.

In that regard, we are working on a novel approach that utilizes machine learning in order to allow for larger and faster simulations. In this talk, I will provide some background about the systems of interest and conventional PIC simulations, and then I will introduce our approach and share some preliminary results.