Multi-Physics Metasurface Inverse Design With Cross-Domain Invariant Feature

Abstract

Deep learning has recently reshaped the landscape of metasurface inverse design by creating a simulation agent from electromagnetic response to structural configuration. Despite significant progress in transfer learning, the inverse design across different physical fields remains challenging due to substantial domain discrepancies. Here, we propose a transfer-learning-assisted inverse design framework that leverages multiple-kernel maximum mean discrepancy (MK-MMD) to bridge the distribution gap between electromagnetic and acoustic fields. By enforcing feature alignment through MK-MMD, the model learns domain-invariant representations, effectively mitigating the design space mismatch between electromagnetic and acoustic metasurfaces. Using electromagnetic metasurfaces as the source domain, our approach successfully transfers knowledge to acoustic metasurface design, enabling high-precision phase-modulation control with merely 150 target-domain samples. Compared to existing methods, our framework reduces the mean squared error (MSE) by over 50% and lowers the data requirements by more than 40%. Our work establishes a sustainable and efficient inverse design framework for multi-physics metasurfaces, paving the way for intelligent adaptive cross-physical-field meta-devices.