Machine-learning-based inverse design of wide-band metasurfaces with interdependent unit cells

Artikel i vetenskaplig tidskrift, 2026

Metasurfaces are thin optical materials consisting of tiny meta-atoms that locally change the phase, amplitude, and polarization of light. Such metasurfaces have been shown to exhibit optical responses beyond what is possible with natural materials, e.g. in thin achromatic lenses, holograms, reflectors, and the generation of structured light. A considerable challenge lies in designing the sometimes thousands of meta-atoms that are needed to locally modify the wavefronts of light. Here we use a conditional generative adversarial network to demonstrate for the first time a way to design metasurfaces where the scattering responses of different meta-atoms are interdependent, i.e. some of the individual scattering parameters are unimportant, and it is the relative values between different meta-atoms that matter. We illustrate our approach with a design of a broadband phase mask for a Zernike wavefront sensor, with complex requirements on the phase relationship between the scattering parameters of an inner disk and the surrounding region. Our inverse design method offers increased design freedom, since it does not require specifying absolute scattering parameters, and allows for metasurfaces with large bandwidths.