Neural networks for the analysis of traced particles in kinetic plasma simulations

Cosmic-ray acceleration processes in astrophysical plasmas are often investigated with fully kinetic or hybrid kinetic numerical simulations, which enable us to describe a detailed microphysics of particle energization mechanisms. Tracing of individual particles in such simulations is especially useful in this regard. However, visually inspecting particle trajectories introduces a significant amount of bias and uncertainty, making it challenging to pinpoint specific acceleration mechanisms. Here, we present a novel approach utilizing neural networks to assist in the analysis of individual particle data. We demonstrate the effectiveness of this approach using the dataset from our recent particle-in-cell (PIC) simulations of non-relativistic perpendicular shocks, which consists of 252 000 electrons, each characterized by their position, momentum, and electromagnetic field at particle’s position, recorded in a time series of 1200 time steps. These electrons cross a region affected by the electrostatic Buneman instability, and a small percentage of them attain high energies. We perform classification, regression, and anomaly detection algorithms on the dataset by using a convolutional neural network, a multi-layer perceptron, and an autoencoder. Despite the noisy and imbalanced dataset, all methods demonstrate the capability to differentiate between thermal and accelerated electrons with remarkable accuracy. The proposed methodology may considerably simplify particle classification in large-scale PIC and hybrid simulations.