Physics-Informed Neural Networks for Multiphysics Simulations: Application to Coupled Electromagnetic-Thermal Modeling

We demonstrate a new approach to building multiphysics solvers, employing physics-informed neural networks (PINNs), through the example of coupled electromagnetic-thermal simulations. In this example, the well-known Finite-Difference Time-Domain (FDTD) method for electromagnetic field simulation is combined with a PINN, designed to replace a thermal solver. The PINN is trained in an unsupervised manner by implementing the heat equation and boundary conditions into the network. As a result, the cost of generating "ground truth" data is eliminated. Our work enables standalone electromagnetic simulators, like FDTD, to solve multiphysics problems accurately and efficiently.