Data-Driven Inverse Design of Hybrid Waveguide Gratings Using Reflection Spectra via Tandem Networks and Conditional VAEs

This study presents a data-driven inverse design approach for one-dimensional hybrid waveguide gratings using full reflection spectra across the visible range and a complete span of incident angles. Traditionally, designing such structures to achieve specific optical responses relies on parameter sweeps and iterative simulations which are computationally expensive, time-consuming, and often inefficient. To overcome this, we generated a comprehensive dataset using rigorous coupled-wave analysis (RCWA) simulations and trained two machine learning models: a deterministic tandem network and a generative conditional Variational Autoencoder (cVAE). Both models were trained on noisy reflection spectra to mimic real-world measurements. They both predict structural parameters accurately on clean and noisy data. On clean data, the mean absolute error (MAE) for silver thickness and grating period is below 1 nm. For the dielectric layer, the error is about 13–15 nm. When noise is added, the Tandem network performs best with low to moderate noise. The cVAE, however, stays more stable under high noise conditions. At σ=0.3, the cVAE model reliably predicts the silver thickness and grating period, with MAEs below 6 nm. The main error comes from the dielectric thickness. Sensitivity analysis of reflection spectra confirms this trend. The reflection is least sensitive to the dielectric thickness, while silver thickness and grating period dominate. This analysis provides physical insight for waveguide design as well in which, accurate control of silver thickness and grating period is far more critical than small errors in dielectric thickness. In general, our approach enables rapid prediction of structural parameters of hybrid waveguide gratings from reflection spectra. This reduces design time and reliance on complex microscopic measurements, with potential applications in sensing, communication, and integrated photonics.