GC-500K: A Physics-Consistent Synthetic Dataset for Photonic Grating Coupler Design

GC-500K is a large-scale, physics-consistent synthetic dataset for photonic grating coupler analysis and data-driven modeling. The dataset contains 500,000 independently generated grating coupler configurations, each represented by wavelength-resolved reflectance (R), transmittance (T), and absorbance (A) spectra sampled at 100 uniformly spaced wavelength points spanning 1200–1600 nm. In addition to spectral responses, the dataset includes five geometric design parameters per sample and a set of derived physical performance metrics.

All samples are generated using a physics-consistent synthetic data generation framework in which fundamental physical constraints are enforced at the data level. In particular, absorbance is computed via the energy-conserving closure relationA = 1 − R − T,ensuring passivity and energy conservation for every wavelength sample. Geometric parameters are sampled independently from uniform distributions over physically admissible fabrication ranges, enabling unbiased coverage of the design space and preventing implicit parameter correlations.

The dataset is accompanied by a comprehensive statistical validation, including tests for parameter independence, geometry–physics consistency, spectral smoothness, numerical stability, output-space coverage, and train–test equivalence. These validations confirm that GC-500K is free from degenerate structure and suitable for downstream use in forward modeling, inverse design, surrogate modeling, and representation learning tasks in nanophotonics.

GC-500K is provided as a compressed HDF5 archive for efficient storage and partial loading. The dataset is intended for research, benchmarking, and educational use and does not represent experimental measurements. Any future extensions or modifications will be released as new dataset versions to ensure full reproducibility.

Intended Use

Machine learning for photonic forward and inverse design

Statistical analysis of geometry–spectral relationships

Surrogate modeling and representation learning

Benchmarking physics-consistent synthetic data methods