Performance evaluation of Parallel PINN based on Domain Decomposition Method

Physics-informed neural network (PINN) is a machine learning method that has been proposed as a new partial differential equation solver and has attracted much attention as an alternative to the finite element method (FEM). Domain decomposition method (DDM) is known as a parallel numerical method for the FEM, in which the analysis domain is divided into smaller subdomains and solution in the whole domain is obtained by solving the problem in each subdomain. In this research, a PINN program for four-domain problems, which is based on the classical DDM algorithm described by partial differential equations (DDM-PINN), is developed, and performance is evaluated with respect to parallel computation and prediction accuracy.