Research On Design Of Metasurface Structures Based On Deep Learning

Metasurface structures have great advantages in controlling optical responses,and can efficiently adjust the phase size,polarization angle,and propagation mode of electromagnetic waves during transmission.They have been widely used in beam steering,super-resolution imaging,nonlinear optics,and other fields.The metasurface structures are usually designed by the full-wave numerical simulation method.However,when the structure is complex,a large number of modeling,parameter sweeping and optimization processes are needed,and the requirements for human experience,computational time and resources are relatively high.In recent years,the developing of artificial intelligence algorithms has accelerated the design process of metasurface,which do not require complex modeling and parameter sweeping processes and can improve the design quality of structures,leading to very important research significance and application value.This paper mainly studies the design of metasurface structures based on deep learning.The main work includes:(1)An attention-based Convolutional Neural Networks-Long Short-Term Memory(A-CNN-LSTM)model is introduced for forward prediction of metasurface spectra.The model is based on a convolutional neural network,on which two neural networks,long short-term memory and attention mechanism,are introduced to improve the performance of the entire network model.In the process of making the data set,the script and cutting and denoising preprocessing of automatic software simulation are used to ensure the quality of the data set.Experiments show that the network model can realize the spectral prediction of metasurface structures of different shapes within few times,and the results are basically consistent with the spectral results obtained by FDTD Solutions simulation software.Its accuracy(up to 98%)is better than CNN and CNN-LSTM.(2)A conditional deep convolutional generative adversarial network(Self-AttentionConditional-Deep Convolutional Generative Adversarial Networks,SA-C-DCGAN)model is proposed to realize the reverse design of the metasurface structures.The model uses deconvolution and convolution operations in the generator and discriminator parts respectively,and then generates the image part by introducing conditional information to enable the network to generate a structure that conforms to the label information.Finally,a self-attention mechanism is added between the last two layers of the convolution network.It is used to improve the effectiveness of generated images.Through a large number of experiments and compared with other traditional generative network models,the high accuracy of the generated images(up to 95%)of the network model is verified.In addition,the computer vision OpenCV method is further studied to extract the geometric parameters of the above-designed metasurface structure,so as to extend the inverse design model to practical applications.This paper provides an efficient,high-precision,and automated design scheme for the design of metasurface structures based on deep learning,which can provide theoretical guidance for the design and application of nanophotonic device structures.