摘要
Conventional transmissive metasurface design for Internet of things(IoT)communications relies heavily on expert knowledge and iterative full-wave simulations,resulting in high computational cost and limited efficiency.To address this challenge,we propose an end-to-end deep-learning-based design framework for sub-6 GHz transmissive communication metasurfaces,which enables automatic mapping from target transmission responses to manufacturable physical structures.We first develop a prediction model,Img2S,to accurately estimate the S-parameters of metasurfaces,significantly reducing the need for full-wave simulations.Based on this model,two variational generative networks,strictly constrained-conditional variational autoencoder(SCCVAE)and loosely constrained-conditional variational autoencoder(LC-CVAE),are proposed to synthesize physically realizable metasurface structures by incorporating geometric priors and electromagnetic consistency constraints.Experimental results show that Img2S achieves a mean squared error(MSE)of 9.76×10^(-4)in predicting the simulated S-parameters of metasurfaces over the operating frequency band.Both simulation and measurement results confirm that the generated metasurfaces closely match the target electromagnetic responses,with single-state mean absolute errors(MAEs)below 0.16 in simulation and below 0.31 in measurement,respectively,outperforming conventional design approaches in terms of accuracy and frequency stability while significantly improving the overall design efficiency.
基金
supported by the National Natural Science Foundation of China under Grant 92567202。
