Machine Learning Multi-Scale Plasma Chemistry

Cold atmospheric plasma (CAP) refers to a non-thermal plasma encountered in ambient air, with diverse applications in cancer therapy, wound treatment, sterilization, agriculture, and air and water purification. These applications critically depend on CAP chemistry, specifically the presence of reactive oxygen and nitrogen species. However, the comprehensive measurement of the numerous species involved in CAP chemistry, which involves complex dynamic chemical reactions, poses a significant challenge for traditional experimental approaches. Additionally, numerical simulations encounter computational costs due to the vast disparity in timescales between sub-nanosecond inelastic collisions (chemical reactions) and the millisecond to minute timescales of CAP operation. In this study, we propose a novel approach that combines physics-based data-driven modeling and advanced machine-learning techniques to overcome these challenges. Our methodology employs a physics-informed neural network (PINN) that is trained using constraints derived from physical laws and experimental measurements. The trained PINN provides a holistic understanding of species concentrations, aligning with physical laws and real-world observations. By integrating experimental measurements with machine learning numerical simulations, our approach presents a general methodology to address the multiscale nature of microscopic plasma chemistry coupled with macroscopic gas flows.