Inverse design of discrete pixelated metasurfaces via masked spectrum-guided diffusion

In metasurface design, two categories of practical design scenarios frequently arise: a designer may specify an idealized spectral target—such as a linear phase dispersion for an achromatic device—that does not correspond to any structure in the simulation database, or may possess only partial spectral observations and need to reconstruct the complete broadband response alongside a corresponding physical structure. We propose a unified masked spectrum-guided inverse design framework that addresses both scenarios within a single, continuous denoising dynamic, in which dynamics masking serves as the key training mechanism. For idealized targets, the framework generates discrete pixelated binary structures whose simulated responses closely approximate any prescribed target profile within the model’s 2–18 GHz training band. For partial observations, the framework simultaneously reconstructs the complete broadband spectrum and the corresponding physical structure. To enable end-to-end surrogate-assisted spectral optimization for discrete binary metasurfaces, a pre-trained forward electromagnetic surrogate model combined with a Straight-Through Estimator (STE) is incorporated into the training loop to sustain continuous spectral gradient supervision through the non-differentiable binarization step. The framework is validated through full-wave simulation and experimental far-field measurement of a representative anomalous reflector operating over a 9–11 GHz sub-band.