Physics-based Deep Learning for electromagnetics

Numerical simulations of electromagnetic wave propagation problems primarily rely on spatially and temporally discretization of the system of time-domain Maxwell equations using finite difference or finite element type methods. For complex and realistic three-dimensional situations, such a process can be computationally prohibitive, especially in view of many-query analyses (e.g., optimization design and uncertainty quantification). Therefore, developing cost-effective surrogate models is of great practical significance. Among the different possible approaches for building a surrogate model of a given PDE system in a non-intrusive way (i.e., with minimal modifications to an existing discretization-based simulation methodology), approaches based on neural networks and Deep Learning (DL) has recently shown new promises due to their capability of handling nonlinear or/and high dimensional problems. In the present study, we propose to focus on the particular case of Physics-Informed Neural Networks (PINNs). PINNs are neural networks trained to solve supervised learning tasks while respecting any given laws of physics described by a general (possibly nonlinear) PDE system. They seamlessly integrate the information from both the measurements and partial differential equations (PDEs) by embedding the PDEs into the loss function of a neural network using automatic differentiation.

Scattering of a plane wave by a dielectric disk: comparison between exact and predicted solution. In the proposed approach, we train two neural networks (one for each medium) so as to include the transmission condition between the two media in the loss definitiones, The two networks share an identical MLP with three hidden layers and 50 neurons in each layer. We use sin as the activation function, and a combination of Adam then L-BFGS optimizations (500 iterations for Adam followed by 15000 iterations for L-BFGS).