Deep Learning of Reflection Phase Predection for Arbitrary Coding Metasurface Atoms

Abstract

Digital coding representations of meta-atoms make it possible to integrate with deep learning algorithms to realize intelligent designs of metasurfaces. In the microscope level, each meta-atom consists of 16×16 square sub-blocks covered with and without metal, denoted by microscope coding ‘1’ and ‘0’, respectively. The reflection phase of meta-atom is governed by the microscope coding pattern. Considering two-fold symmetry, there are 264 different coding patterns, corresponding to different phase responses. In this paper, we propose a deep learning method to predict the reflection phase of meta-atoms with arbitrary patterns, at 10GHz, in which only 70000 training coding patterns are used to train the network. We employ the other 10000 randomly-chosen coding patterns to validate the neural network, showing an accuracy of 90.05% of phase responses with 2° error in 360° phase and the average forecast error is 1.4933°. Using the learned network, we can readily find the correct coding pattern among 18 billion of billions of choices for required phase in a second, finishing fast automatic design of arbitrarily anisotropic meta-atoms. If the traditional method is used, the design of arbitrarily anisotropic meta-atoms would be very complicated and time consuming, in which experts must find special geometries based on large amounts numerical simulations to fulfill the tasks.