Spectrum prediction and inverse design of metasurfaces via transfer learning based on material similarity

Spectrum prediction and inverse design of metasurfaces based on deep learning have been a hot research topic. The dependence of deep learning on data is a major challenge for its widespread application in the field of metasurfaces. In this letter, we proposed a transfer learning method based on material similarity to accomplish spectrum prediction and the inverse design of metasurfaces. As a proof-of-concept, we investigated the transfer tasks of two types of metasurface, i.e., absorption metasurface and polarization conversion metasurface, whose material properties could be represented by the Drude model to reflect the material similarity, and accomplished the spectrum prediction and inverse design through transfer learning. We achieved 50% data saving, demonstrating reduction of the reliance on training data volume while ensuring network performance. The proposed concept may provide a new avenue for metasurface and metamaterial designs.