Electromagnetic-Thermal Cosimulation via Divide-and-Conquer PINNs

A physics-informed neural networks (PINN) framework is proposed for electromagnetic-thermal co-simulation. The framework utilizes two coupled PINNs: the first solves the Helmholtz equation for electric field distribution, while the second predicts temperature distribution through the heat conduction equation with electric field-derived heat sources. Regarding the complexities of multi-material regions with substantial property contrasts in heterogeneous media, a divide-and-conquer strategy is devised to enhance the accuracy of the PINNs. Numerical experiments on perfectly electrically conducting (PEC) cavities demonstrate that the trained PINN-based electromagnetic-thermal co-simulation method significantly reduces both computational time and memory usage when compared to conventional finite element method (FEM) simulations, while preserving high-accuracy predictions.