A general framework for knowledge integration in machine learning for electromagnetic scattering using quasinormal modes

Abstract

Neural networks have been demonstrated to be able to accelerate the modeling and inverse design of optical and electromagnetic devices by serving as fast surrogates for electromagnetic solvers. Nevertheless, such neural networks can be unreliable and normally require extreme amounts of data to train. Here it is shown that these limitations can be alleviated by constraining neural-network models using prior knowledge about the governing physics. We propose a universal physics-informed neural network framework for electromagnetic scattering based on the quasinormal mode expansion of the scattering matrix. The neural networks learn the resonant structure underlying the scattering spectrum, are guaranteed to obey energy conservation and causality, and are shown to have significantly improved data efficiency for photonic-crystal slabs and all-dielectric free-form metasurfaces. Furthermore, the framework allows additional problem-specific constraints, such as losslessness, symmetries, and number of modes, to be imposed manually when they are available. The method can be applied to a wide range of optical and electromagnetic devices owing to the generality of the quasinormal mode formalism.