Numerical simulations on DC-driven low-temperature plasmas by physics-informed neural networks

Abstract

In recent years, artificial intelligence (AI) technology is empowering various fields, and the combination of AI and numerical simulation of Low-Temperature Plasmas (LTPs) has attracted widespread attention. In this study, the physics-informed deep neural networks (PINNs) are used to solve the fluid model, which describes a one-dimensional DC-driven discharge system. PINNs use a multilayer perceptron to represent the solutions to the poisson equation, continuity equation, and drift-diffusion approximation equations in fluid model and employ automatic differentiation to compute the derivatives of these field variables, thereby constructing the residuals of the physical equations. The training dataset consists of randomly sampled spatial and temporal coordinates along with their corresponding distribution functions, generated from fluid simulations. The gradient descent algorithm minimizes the sum of the residuals of the physical equations and the data loss to update the model parameters, enabling PINNs to fit observed data while satisfying the fluid equations. The simulation results show that this mesh-less method can effectively solve the fluid model at any spatial-temporal resolution instead of discretization method with a fixed spatial-temporal resolution, providing an alternative for numerical solution of fluid models for LTP in the era of AI.