Integrating Deep Learning with 2D-3V Particle-in-Cell (PIC) simulations of Low Temperature Plasmas (LTP)

PIC-MCC method is a widely used technique for LTP simulation, however, device-scale 2D/3D PIC simulations are computationally intensive for capturing multiscale physical phenomena, necessitating parallel implementations [1]. A major parallel performance bottleneck arises in the Charge Deposition (CD) module, which interpolates particle charges onto the grid, which is later used to calculate potentials and electric fields. To address this limitation, we propose a hybrid data-driven approach that can accelerate PIC simulations by replacing the traditional CD, Poisson Solver, and Electric Field computation with a deep learning-based model. We employ a U-Net architecture to learn from a 2D histogram of charged particles to predict the electric fields and further enhance physical fidelity via Physics-Informed Neural Networks. This proof of concept model is being tested for an ExB based plasma simulation under multiple test cases. After running the simulation for 500 iterations for a 128x128 grid with 2.6x10⁵ particles, the proposed hybrid approach achieves a mean percentage error of less than 5% compared to a standard PIC-MCC simulation, demonstrating its potential as an efficient and accurate alternative for PIC-based plasma simulations.

[1] Investigation of EDF evolution and charged particle transport in E × B plasma based negative ion sources using kinetic simulations. Sci Rep 13, 20044 (2023).