Comparison of Deep Learning based approaches for Simulation of Microwave-Plasma interaction

Recent efforts to explore UNet based Deep Learning (DL) approach for 2D simulations of complex microwave–plasma interactions have shown promising results in terms of speed [1]. However, enhancement in accuracy is still required for its widespread application. Investigating this interaction is particularly challenging when dealing with inhomogeneous plasma. The DL model must learn the complex interplay of transmission, reflection, and absorption of an electromagnetic (EM) wave as it propagates through an inhomogeneous plasma, resulting in intricate EM scattering patterns. In this study, we review four deep learning models, U-Net, Variational Autoencoder, Pix2Pix, and a custom Hybrid CNN-Vision Transformer model. These models are trained on an extensive and diverse dataset covering a wide range of plasma densities, topologies and electric field strengths. Training data have been generated using 2D FDTD-based fluid simulations of microwave breakdown. The trained DL models are then employed to reproduce 2D scattered electric field values for incident microwaves on various plasma density profiles. We compared the results of different DL approaches with physical data obtained from established 2D FDTD-based methods. We find that Hybrid CNN-ViT model performs best on unseen data while delivering a speedup of approximately 100x compared to FDTD. We also support our findings with latent space visualization.

[1] Desai et al 2022 IEEE Transactions on Microwave Theory and Techniques 70 (12) 5359.