CNN‐Transformer Hybrid Network for Efficient Performance Prediction of Metasurfaces

This paper proposes an efficient metasurface simulation method leveraging deep neural networks to mitigate the time‐consuming and resource‐intensive nature inherent in traditional numerical simulation techniques. The proposed approach harnesses convolutional neural networks (CNNs) to extract metasurface features and employs a transformer network for electromagnetic field prediction. To ensure precision, simulation data from the finite‐difference time‐domain (FDTD) method with coarse grid is incorporated as additional input to the transformer. By integrating CNNs and transformers with coarse mesh compensation data, the accuracy of optical response prediction is enhanced. Prediction outcomes demonstrate that the proposed method can predict the electromagnetic characteristics of the metasurfaces accurately and efficiently. The mean squared error is merely 1.35e‐4. Furthermore, this method yields spectral results of the metasurface within a mere 3 s, marking a notable 100‐fold increase in efficiency compared with traditional simulation software.