線形静磁場解析におけるPINNsの並列計算

抄録

Physics-informed neural network (PINNs or PINN) is supervised learning for approximating the initial-boundary value problems of partial differential equations. This research focuses on the magnetostatic field problem derived from Maxwell's equations and evaluate the performance of large-scale computation and parallel computation of PINN. In finite element analysis, which is a conventional numerical analysis method, the magnetostatic field is formulated with the magnetic vector potential ‘A’ as an unknown. However, since the three-dimensional magnetostatic field problem has indefiniteness, it is necessary to solve a singular problem or construct an equation with the indefiniteness removed. This study also applies PINN to partial differential equations based on the A-formulation.