Deep learning-based inverse design of broadband metasurface polarization converter

Deep learning offers promising new methods for the design and optimization of electromagnetic metasurfaces. In this study, we propose a hybrid deep learning model that integrates a Convolutional Neural Network (CNN) with a Transformer architecture (CNN-Transformer) for the inverse design of broadband polarization converters, using polarization conversion ratio spectra as input. To enhance the robustness and generalization capability of the model under limited data conditions, Gaussian noise is incorporated into the training data as a data augmentation strategy within a few-shot learning framework. In comparison with a conventional Multilayer Perceptron (MLP), the CNN-Transformer model achieves markedly superior performance, attaining an average mean squared error of 0.003 31 and significantly outperforming the MLP baseline. Utilizing the structural parameters predicted by the model, a broadband linear polarization converter was designed and fabricated, demonstrating excellent polarization conversion efficiency across the 4–16 GHz frequency range. Both the simulation and experimental results confirmed the accuracy and effectiveness of the proposed method for structural parameter prediction. This study substantially streamlines the design process of electromagnetic metasurfaces, delivers high design precision and robust performance, and further highlights the potential of deep learning in advancing the development of complex metasurface devices.