Differentiable generative design of multi-wavelength Si-on-glass metagratings

Multi-wavelength dielectric metagratings require subwavelength material patterns that balance wavelength-dependent diffraction, material dispersion, and binarization. We combine the GLOnet generative inverse-design architecture with FMMAX, an open-source JAX-native rigorous coupled-wave analysis (RCWA) library, to optimize this problem with end-to-end automatic differentiation. The revised physical protocol uses one fixed 1039.23 nm period and targets the transmitted +1 diffraction order at 700, 800, and 900 nm. The corresponding air-side output angles are 42.34 ∘ , 50.34 ∘ , and 60.00 ∘ , respectively, as required by the grating equation. This defines one physical silicon-ridge metagrating rather than three wavelength-scaled geometries. The framework produces distributions of binary silicon/air candidates and supports a multi-wavelength objective without wavelength-specific adjoint derivations. Across five independent GLOnet training seeds, the mean three-wavelength candidate efficiency was 0.625 ± 0.091 (mean and sample standard deviation of the five seed-level means). A matched 500-start Adam direct pixel-optimization baseline had a final-iterate ensemble mean of 0.639; retaining the best binary checkpoint along each trajectory gave 0.699 ± 0.100. The best candidate from the best-of-five GLOnet run reached a mean of 0.884 and a three-wavelength bottleneck of 0.843, compared with direct-baseline maxima of 0.835 and 0.798. Thus, the generator exhibited a higher best-case ceiling but substantial seed dependence, whereas direct optimization was the stronger typical baseline. These results establish a reproducible differentiable generative workflow while delimiting, rather than assuming, its optimization advantage.