Research On Metasurface Design Based On Deep Learning

Metasurfaces,as a new concept of artificial electromagnetic materials with unique electromagnetic wave manipulation capabilities,have been widely applied in various fields.However,traditional metasurface design methods rely on extensive computational resources and require experienced professionals to spend a significant amount of time on design.With the rapid development of artificial intelligence,deep learning techniques provide a new approach for convenient metasurface design.In this paper,we study the establishment of the correspondence between metasurface structural parameters and their electromagnetic responses using deep learning techniques,and based on this,we achieve fast design of two common types of metasurfaces: continuous parameter structure and digital coding structure.The proposed methods show significant improvements in design accuracy and speed compared to existing literature methods.The main research contents are as follows:Research on deep learning-based design methods for continuous parameter structure metasurfaces.Considering the characteristics of the electromagnetic transmission response curves of continuous parameter structure metasurfaces,we propose two design methods: one-dimensional convolutional neural network and Gram-Attention-Scene-Field convolutional neural network(GASF-CNN).The former directly extracts features from the response curves during the design process,while the latter transforms one-dimensional data into two-dimensional images through Gram-Attention-Scene-Field and then extracts features.The results of designing dualband metasurfaces as design examples show that both models can automatically optimize the structural parameters based on the input desired bandgap curve,and demonstrate high design efficiency,model accuracy,and generalization ability.Compared with traditional machine learning methods,they have significant advantages in evaluation metrics such as mean squared error(MSE)and mean absolute error(MAE).(2)Research on digital coding structure metasurface design method based on transfer learning.Due to the limitations of prior unit structure patterns in the design flexibility of continuous parameter structure metasurfaces,and the increasing popularity of digital coding metasurfaces,we propose a transfer learning and attention mechanism-based design method for digital coding metasurfaces.By leveraging pretrained convolutional neural network models and bilinear interpolation techniques in a transfer learning scheme,we achieve fast prediction of the reflection phase of 210x10 metasurface unit cells.To further improve the performance of the prediction model,we also introduce attention mechanism techniques to help the model focus more on the most relevant parts of the input data by assigning different weights to different regions.The results show that the accuracy of the transfer learning-based model reaches over 90%,and the improved model with attention mechanism achieves an accuracy of over 95%,demonstrating significant performance advantages compared to other networks.To validate the design capability of the proposed method,we use the obtained trained model to design a set of beam-steering gradient phase metasurfaces,showing great potential in relevant applications.