Consistency‐Constrained Domain Adaptation for Data‐Efficient Inverse Design of Metasurfaces

ABSTRACT Metasurfaces enable precise light‐field manipulation at subwavelength scales, but deep learning‐driven inverse design remains limited by domain shifts arising from changes in materials, structural topologies, and fabrication conditions. These shifts often require costly target‐domain data acquisition and model retraining. This work identifies a failure mode in cross‐domain inverse design, termed the “pseudo‐solution trap” in which standard domain adaptation improves feature alignment but still yields inverse solutions whose reconstructed responses deviate from the targets. To address this issue, a consistency‐constrained domain adaptation framework is proposed for data‐efficient metasurface inverse design. The method uses a calibrated forward surrogate to impose a forward‐consistency constraint during inverse‐model adaptation under domain shift. A small labeled target‐domain set is used to calibrate the surrogate, which then regularizes adaptation on unlabeled target‐domain data. This simulation‐grounded strategy reduces physically inconsistent solutions during cross‐domain transfer. In few‐shot supervised settings, high‐precision inverse design is achieved with as few as 20 labeled samples, while in fully unlabeled target domains the framework improves robustness relative to standard unsupervised domain adaptation. Numerical validation on metasurface‐based optical filter design shows improved inverse‐design fidelity under representative topology and material/frequency shifts.