Semisupervised Generative Model for Design of Polarization Conversion Metasurfaces

This paper proposes a semisupervised generative model for unit cell sample generation to design polarization conversion metasurfaces (PCMs). This model obtains new unit cell structures and corresponding electromagnetic (EM) responses from simple probability distributions in the latent space to achieve better results. The semisupervised scheme combines supervised learning and unsupervised learning by sharing their respective weights of the encoders and decoders, reducing the number of labeled samples required to half the total number of samples. To improve the design freedom of unit cell structure, nonparametric modeling is introduced with binary images. To verify the proposed semisupervised generation model, a quasi‐I–shaped pattern is selected as the basic unit cell structure for validation. After the trained generative model generates the optimal unit cell, a chessboard ultrawideband PCM with dimensions of 240 × 240 mm is designed, and the monostatic radar cross section is reduced by 10 dB from 10.2 to 18 GHz under the far‐field normal incidence. The proposed model expands the freedom of the unit cell structure and improves the PCM performance using only a small number of labeled samples. Furthermore, compared with traditional generative models using the same training set, it shows significant improvements in both unit cell structure image generation and EM response prediction. This research provides an efficient new method for the intelligent design of microwave metasurfaces and supports their subsequent engineering applications.