Graph Neural Networks for Metasurface Modeling

Abstract

When using deep neural networks to model electromagnetic fields, one often needs to fix spatial sizes of problems to fit the input dimension of neural networks, which is determined during the training process. This limitation makes it difficult to use neural networks to model different metasurfaces with varying sizes, particularly when there is strong coupling between the scattering units in the metasurface. We propose a Graph Neural Networks (GNN) architecture which learns to model electromagnetic scattering, and it can be applied to metasurfaces of arbitrary sizes. Most importantly, it takes into account the coupling between scatterers. Using this approach, near-fields of metasurfaces with dimensions spanning hundreds of times the wavelength can be obtained in seconds. Our approach can also be used for the inverse design of large metasurfaces.