Prediction and Anomaly Detection of accelerated particles in PIC simulations using neural networks

TL;DR

This paper introduces neural network techniques to analyze particle trajectories in PIC simulations, enabling accurate energy prediction, classification, and anomaly detection, thereby improving understanding of particle acceleration in astrophysical plasmas.

Contribution

The paper presents a novel neural network-based approach for analyzing particle data in PIC simulations, enhancing classification, regression, and anomaly detection capabilities.

Findings

Neural networks accurately predict particle energies despite noisy data.

Classification and anomaly detection effectively distinguish energetic particles.

Method simplifies analysis of large-scale PIC simulation data.

Abstract

Acceleration processes that occur in astrophysical plasmas produce cosmic rays that are observed on Earth. To study particle acceleration, fully-kinetic particle-in-cell (PIC) simulations are often used as they can unveil the microphysics of energization processes. Tracing of individual particles in PIC simulations is particularly useful in this regard. However, by-eye inspection of particle trajectories includes a high level of bias and uncertainty in pinpointing specific acceleration mechanisms that affect particles. Here we present a new approach that uses neural networks to aid individual particle data analysis. We demonstrate this approach on the test data that consists of 252,000 electrons which have been traced in a PIC simulation of a non-relativistic high Mach number perpendicular shock, in which we observe the two-stream electrostatic Buneman instability to pre-accelerate a portion of electrons to nonthermal energies. We perform classification, regression and anomaly detection by using a Convolutional Neural Network. We show that regardless of how noisy and imbalanced the datasets are, the regression and classification are able to predict the final energies of particles with high accuracy, whereas anomaly detection is able to discern between energetic and non-energetic particles. The methodology proposed may considerably simplify particle classification in large-scale PIC and also hybrid kinetic simulations.