Domain adaptation based generative model for polarization conversion metasurface design

In this paper, a domain adaptation (DA) assisted semi-supervised generative model is proposed for polarization conversion metasurface (PCM) design. With the help of unsupervised learning, the semi-supervised generative model reduces the number of required labeled samples and lowers the difficulty of training. DA that is achieved by constructing an input migration and an output migration is used for transfer learning. The proposed model provides units for an ultra-wideband low radar cross section (RCS) PCM. The designed PCM using units with seven opening angles maintains a 10 dB monostatic RCS reduction from 10.2 to 18 GHz compared with a metal plate. Furthermore, the designed PCM achieves a 7 dB reduction in the maximum bistatic RCS at the center frequency compared with a PCM using units with a single opening angle. The introduction of DA enables topological modeling of units of the same type but with different features based on the trained generative model, improving the design efficiency.