Design Of Chirality Metasurface Devices Based On Deep Learning

Metasurface materials are a special kind of artificial layered structure,whose thickness is thinner than the wavelength.It can effectively adjust the polarization method,amplitude,phase,polarization,and transmission model of electromagnetic waves,thereby achieving accurate control of electromagnetic waves.Chiral materials commonly exist in nature,such as a variety of amino acids and proteins,and have the characteristic that the configuration of the material cannot coincide with its mirror enantiomers.In chiral materials,there is an asymmetric transmission effect of linearly polarized light,which can effectively improve the performance of optical devices and regulate polarization.In addition,it can also reduce the volume of the device at the same time.However,due to traditional design methods such as FEM and FDTD,the design process of chiral metasurfaces is very cumbersome and time-consuming.This thesis aims to improve the efficiency of design by training a deep learning neural network and using it to design chiral metasurfaces for polarization control.In the selection of chiral metasurfaces,a two-layer continuous U-shaped structure was selected.In the deep learning neural network,the generation confrontation network is selected.The network can output its corresponding three structural parameters by inputting 51 asymmetric transmission coefficients representing polarization regulation capabilities from 2.5THz to 3.5THz to the network.After completing the design training of the network,this thesis tested the performance of the network and conducted a reverse design of the two-layer continuous U-shaped structure.The thickness of the substrate,the length of the linear portion of the element structure,and the cycle length of the element are selected as the output of the network.51 asymmetric transmission coefficients of 2.5THz to 3.5THz are used as input to the network.The results show that after using the three output structural parameters for simulation,the generated asymmetric transmission coefficient curve can well fit the corresponding curve of the real sample,reducing the need for professional ability of the designer,bypassing a large number of calculation processes,and providing assistance for designing chiral metasurfaces for polarization control more quickly and effectively.