Preventing Structural Fragmentation in Metasurface Inverse Design via 2D Fourier Series and Boundary Optimization

Discretized metasurfaces offer freeform flexibility in electromagnetic response manipulation, but higher structural resolution causes random generation to produce fragmented pixelated geometries that are hard to fabricate and degrade performance. We propose an inverse design framework centered on a Fourier-based metasurface generator that uses 2D Fourier series to maintain geometric diversity while suppressing structural fragmentation. The Kramers–Kronig relation is embedded as a physical prior in the spectrum prediction network to shorten training while preserving accuracy, and a boundary detection mechanism is incorporated into the genetic algorithm to maintain geometric continuity during optimization. The framework is validated on linear-to-circular and linear-to-elliptical polarization conversion tasks, with agreement between simulations and measurements confirming the effectiveness of the proposed approach.