Optimizing the design of neural networks for plasma modelling based on NAS-PINN: 1-D arc simulation as an example

Due to the infeasibility of experiments in most circumstances, plasma simulation is a more general approach to investigating plasma behavior. Physics-Informed Neural Network (PINN) has been a powerful tool for plasma simulation, thanks to its unsupervised and meshless features. However, the design of neural architectures for plasma modelling still demonstrates over-reliance on prior knowledge and experience, which can be imprecise and computationally expensive. Therefore, by incorporating neural architecture search-guided physics-informed neural network (NAS-PINN), we design a framework for plasma modelling which can automatically find the best neural architecture for plasma modelling. NAS-PINN is a neural architecture search algorithm which can complete quick search in a more flexible and larger search space by combining search space relaxation and tensor masks. In this work, we explore the stationary and transient 1-D arc model as an example to find and summarize the patterns of effective neural architectures for PINN-based plasma modelling. Furthermore, we investigate the different input-output frameworks and decoupling methods in transient 1-D arc model and the results can guide the design of PINN-based plasma modelling in more complex conditions.