A dual-branch multiscale cascaded feature-fusion network for the inverse design of EIT metasurfaces

Abstract

Electromagnetically induced transparency (EIT) metasurfaces have great potential in sensing, slow-light devices, optical switching, and filtering. However, their conventional design relies on repeated full-wave simulations, extensive parameter sweeps, and substantial prior experience, making the process time-consuming and computationally expensive. To address this issue, we propose a dual-branch cascaded feature-fusion network (DB-CFNet) for the inverse design of EIT metasurfaces. In this framework, spectra under two polarization states are fed into two parallel branches for independent feature extraction. Multi-scale convolutional modules are employed to capture both local resonance details and global spectral variations, while intermediate feature-fusion blocks enable progressive interaction between the two branches. As a result, DB-CFNet establishes an effective nonlinear mapping from dual-polarization spectral responses to structural parameters. The results show that DB-CFNet achieves parameter prediction in about 10 ms, which is thousands of times faster than finite-difference time-domain simulations, with representative relative spectral errors ranging from 2.066% to 3.849%. The proposed method significantly improves design efficiency and can be extended to the inverse design of a broad class of metasurfaces with complex spectral responses.