Structure parameters optimization of metasurface for broadband polarization conversion based on deep learning and optimization algorithm

Metasurface can be used to manipulate the polarization, amplitude, phase of electromagnetic waves which has been applied in various fields such as holography technology, imaging and sensing. However, traditional metasurface optimization methods such as parameters scanning require a considerable amount of time and computing power and cannot cover every set of parameters due to the setting of step size, which may lead to the optimization result falling into a local optimal solution. To address the above-mentioned issues, in this paper, we propose a method that combines transfer learning optimization network with grey wolf optimization algorithm for structural parameters optimization. Using deep learning networks can significantly improve the speed of spectral prediction, while transfer learning algorithms can enhance the prediction accuracy of the networks. The grey wolf optimization algorithm belongs to a global optimization algorithm and its performance is superior to other traditional algorithms, thus enabling it to achieve a wider bandwidth. The results show that the bandwidth of the transmission spectrum obtained through the grey wolf optimization algorithm is 87.37 nm, which is wider than that achieved through traditional method of parameter scanning. At the same time, it only takes 45 minutes, which is one-seventieth of the time required by traditional optimization methods.

I. Introduction

Currently, the development of metasurface provides an effective shortcut for the breakthrough of applied technology, including holography technology^[1]-[2], imaging^[3], sensing^[4] and polarization control^[5]. Nowadays, most spectral predictions and metasurface design work can be solved by numerical simulation of Maxwell's equations using methods such as the Finite Element Method (FEM) and the Finite Difference Time Domain (FDTD)^[6]. However, these methods often require a considerable amount of simulation time and computing resources. Especially it takes even more efforts to find the optimal performance based on the optical response of metasurfaces. Traditionally, researchers often rely on previous experience and knowledge to find the optimal structural parameters metasurfaces through parameter scanning^[7].