Accelerating Cross‐Scenario Metasurface Adaptability with Plug‐and‐Play Kernel

Abstract The adaptivity of metasurfaces depicts an important behavior of dynamically adjusting their internal properties to adapt to changing environments and cater to various demands. In spite of the proliferation of intelligence‐driven metasurfaces, they still lack the analogical reasoning to execute cross‐scenario tasks due to the random and unpredictable scattering events. Here, a compatible plug‐and‐play scheme is proposed to assist metasurfaces adapt from one to another environment. Unlike conventional step‐by‐step training, such a scheme enables the rapid establishment of reliable neural networks with physical interpretability by integrating past experience into a package kernel. Taking the invisibility cloak as a demonstration, a fully intelligent system is set up, and carried out a three‐stage adaptability validation: point‐to‐point, path‐to‐path, and scene‐to‐scene, suggesting a substantial speed improvement compared with direct learning. Furthermore, by capitalizing on the resonant metasurfaces, the plug‐and‐play kernel is successfully extended to broadband performance with a relative bandwidth of 50%. work provides a powerful and easy‐to‐access recipe for advancing metasurfaces in new and even unknown environments and fosters the shining and bloom of intelligent adaptive meta‐devices.