Electromagnetic Scattering from a Three-dimensional Object using Physics-informed Neural Network

Prediction of electromagnetic fields scattered from objects is of great significance in various fields. Traditional computational electromagnetic solvers, which are mesh-based, are expensive and time-consuming. The deep learning technique becomes an alternative method of the prediction of scattered fields with high efficiency. However, the data-driven deep learning method requires a large data set and lacks robustness. For complicated scattering problems, the construction of a large training data set is a hard task. By considering physics-constraints, physics-informed neural networks (PINNs) can solve the partial differential equation (PDE) problem with a small data set and also provide a physical explanation. In this paper, the PINNs are employed to solve the scattering of a plane wave by a three-dimensional object with Maxwell’s equations being physical constraints. In the calculation, a sphere and an ellipsoid are taken as examples, and the effects of the network parameters (including the number of hidden layers, and the number of data sets) are mainly discussed. The results have practical applications in many fields such as radar detection, biomedical imaging, and satellite navigation.