Reduced Order Residual Neural Networks for Plasma Streamer Propagation

High-fidelity numerical simulations of low-temperature plasmas, such as Particle-in-Cell (PIC) and fluid models, provide critical insights into streamer dynamics but are computationally prohibitive due to the requirement for extreme temporal precision. The coupling of non-linear transport equations, including drift-diffusion and Poisson equations, introduces significant numerical stiffness, often necessitating time steps on the order of $10^{-15}$ seconds. To address these challenges, this study presents a novel reduced-order modeling framework that integrates Deep Residual Networks (ResNets) with autoencoders to simulate plasma streamer propagation efficiently. Utilizing data generated from a one-dimensional fluid solver for argon plasma, the proposed architecture employs an autoencoder to compress high-dimensional spatial state information into a reduced latent space, where a ResNet functions as a time-stepping simulator. A key innovation in this work is the development of a Stiffness-Reduced Autoencoder (SRAE), which incorporates a custom regularization term into the loss function to minimize numerical stiffness within the latent dynamics. Results demonstrate that the hybrid SRAE-ResNet model accurately predicts the temporal evolution of 12 distinct state variables, including electron density and electric field, while substantially reducing the model's parameter space and computational cost compared to full-order simulations. This data-driven approach effectively bridges the gap between predictive accuracy and computational feasibility, paving the way for the integration of efficient phenomenological plasma models into complex computational fluid dynamics (CFD) workflows.