Solving Multi-Scale Plasma Chemistry Using Physics-Informed Neural Network

Cold atmospheric plasma (CAP) is a non-thermal plasma exposed in the open air, which is now applied to cancer therapy, wound treatment, sterilizations, agriculture, air and water purification, etc. All these applications are intensively relying on CAP chemistry, especially the reactive oxygen and nitrogen species. However, due to the complexity of CAP chemistry which contains hundreds of species and thousands of dynamic chemical reactions, it is challenging to measure the full picture of species concentrations through experiments. Numerical simulations are also at high computational cost because of the sub-nanosecond time scale of inelastic collisions (chemical reactions) compared with the millisecond to the minutes time scale of the CAP working period. Therefore, we developed physics-informed data-driven modeling to solve such a multi-scale problem using modern machine learning techniques. The physics-informed neural network (PINN) is trained under the constraints of chemical rate equations with a complete kinetic scheme, conservation laws, and the experimental measurement of a few species concentrations. After the training, PINN can provide us with the full picture of all the species' concentrations which agrees with the physical laws and observations in reality. In other words, an experiment and machine learning-assisted numerical simulation is developed, as a general method that can solve a multi-scale problem of microscopic plasma chemistry coupling with macroscopic gas flows.