Color2Struct: efficient and accurate deep-learning inverse design of structural color with controllable inference

Recent studies demonstrate that deep learning enables inverse design of structural colors by learning the complex nonlinear relations between structural parameters and optical responses. Several models, including tandem neural networks and generative adversarial networks, have been proposed, but they often exhibit bias and scale poorly to more complex structures. Moreover, difficulty incorporating physical constraints during inference hinders the controllability of the output spectra. In this work, we propose Color2Struct, a general framework for efficient and accurate inverse design of structural colors with controllable spectral predictions. Using sampling bias correction, adaptive loss weighting, and physics-guided inference, Color2Struct reduces color difference by 65% and near-infrared reflectance by 48% for sRGB primary colors relative to a baseline tandem-network. We fabricate thin-film nanostructures via standard deposition methods and measure reflectance spectra to validate the model predictions and simulations. The proposed framework can be further generalized to a wide range of research fields beyond nanophotonics.