A Multi‐Resolution Machine Learning Paradigm for Pixelated Metasurface Design

ABSTRACT Pixelated metasurfaces, designed using machine learning, offer a powerful route to advanced optical functionalities. However, achieving high‐resolution designs is limited by significant computational cost and data requirements. This work presents a novel multi‐resolution machine learning framework, coupled with a dedicated optimization algorithm, to overcome these limitations. Our approach leverages complementary information across multiple resolutions to accurately predict electromagnetic responses without the computational burden of high‐resolution‐only designs. By efficiently transferring knowledge between the low‐ and the high‐resolution data, we significantly enhance design flexibility and reduce reliance on extensive datasets. With the multi‐resolution optimization algorithm, the model enables precise control over the amplitude, phase, and polarization of electromagnetic waves, facilitating the efficient design of both reflective and transmissive metasurface devices. We demonstrated the efficacy of this method by designing and experimentally validating broadband polarization converters and polarization‐controlled reflective deflectors. These results provide a practical and scalable framework for creating complex, multi‐functional meta‐devices, unlocking new possibilities for intricate and sophisticated metamaterial design.