Neural network based design of metagratings

Abstract

Metagratings are flat and thin surfaces that rely on unique, periodically repeating (non-gradient), arbitrary shaped light scattering units for wave manipulation. However, the absence of an empirical relationship between the structural and diffraction properties of the units enforces utilization of brute force numerical optimization techniques to determine the unit shape for a desired application. Here, we present an artificial neural network based methodology to develop a fast-paced numerical relationship between the two. We demonstrate the training and the performance of a numerical function, utilizing simulated diffraction efficiencies of a large set of units, that can instantaneously mimic the optical response of any other arbitrary shaped unit of the same class. We validate the performance of the trained neural network on a previously unseen set of test samples and discuss the statistical significance. We then utilize the virtually instantaneous network operations to inverse design the metagrating unit shapes for a desired diffraction efficiency distribution. The proposed inter-disciplinary combination of advanced information processing techniques with Maxwell's equation solvers opens a pathway for the fast-paced prediction of metagrating designs rather than full wave computation.