GC-500K: A Physics-Consistent Large-Scale Benchmark Dataset for Grating Coupler Design

We introduce GC-500K, a physics-consistent syn- thetic dataset of 500,000 grating coupler de- signs for silicon photonics. Each sample com- prises five geometric parameters, 100-point re- flectance, transmittance, and absorbance spectra (1200–1600 nm), derived scalar metrics, and en- ergy conservation flags. The dataset is gener- ated via a validated analytical model with con- trolled noise and is rigorously validated for input independence, output diversity, spectral smooth- ness, and physical plausibility, with fidelity con- firmed against finite-element simulations. A sys- tematic benchmark evaluates 23 models span- ning classical ML, deep learning, operator meth- ods, and generative models across four tasks: forward scalar/spectrum prediction and inverse scalar/spectrum design. Key findings: (i) tree- based models excel at forward scalar prediction (R2=0.978); (ii) UNet1D achieves near-perfect spectral reconstruction (MSE=3×10−5); and (iii) inverse success rates remain < 4%, confirm- ing fundamental non-uniqueness. GC-500K pro- vides a standardized testbed for reproducible data- driven photonic design. The GC-500K dataset and benchmarking codebase will be made pub- licly available upon acceptance.