Neural network enabled metasurface design for phase manipulation

Abstract

The phase of electromagnetic waves can be manipulated and tailored by artificial metasurfaces, which can lead to ultra-compact, high-performance metalens, holographic and imaging devices etc. Usually, nanostructured metasurfaces are associated with a large number of geometric parameters, and the multi-parameter optimization for phase design cannot be possibly achieved by conventional time-consuming simulations. Deep learning tools capable of acquiring the relationship between complex nanostructure geometry and electromagnetic responses are best suited for such challenging task. In this work, by innovations in the training methods, we demonstrate that deep neural network can handle six geometric parameters for accurately predicting the phase value, and for the first time, perform direct inverse design of metasurfaces for on-demand phase requirement. In order to satisfy the achromatic metalens design requirements, we also demonstrate simultaneous phase and group delay prediction for near-zero group delay dispersion. Our results suggest significantly improved design capability of complex metasurfaces with the aid of deep learning tools.