Plasma chemical kinetic simulation based on physics-informed neural networks

Plasma simulation often requires chemical kinetic modeling which describes numerous chemical reactions involving various species. However, due to significant differences in particle concentration scales and time evolution scales, the governing equations of plasma chemical kinetic simulation are usually too stiff to be solved by traditional numerical methods. In this work, we propose a physics-informed neural network (PINN)-based method, i.e. Chem-PINN, for fast and robustly solving stiff equations in plasma chemical kinetic models. Chem-PINN firstly classifies plasma species based on their concentration scales and evolution time scales, using independent neural networks for each class to eliminate stiffness problems caused by differences in species categories. Furthermore, the loss function of Chem-PINN is adjusted during the training by the adaptive weight factor to eliminate the backpropagation gradient imbalance caused by differences in magnitude of each part of the loss function. We tested Chem-PINN with some classical chemical kinetic problems. The results demonstrate that Chem-PINN performs well in solving stiff equations in plasma chemical kinetic modeling. This could provide an alternative approach to plasma chemical kinetic simulation, and also an AI-driven module for plasma fluid simulation.