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

Area of Science:

• Nanophotonics
• Materials Science
• Computational Physics

Background:

• Deep learning models like tandem neural networks and generative adversarial networks are used for inverse design of structural colors.
• These models struggle with bias, scalability to complex structures, and incorporating physical constraints for spectral control.

Purpose of the Study:

• To propose Color2Struct, a general framework for efficient and accurate inverse design of structural colors.
• To achieve controllable spectral predictions by addressing limitations of existing deep learning models.

Main Methods:

• Developed a framework incorporating sampling bias correction, adaptive loss weighting, and physics-guided inference.
• Utilized standard deposition methods for fabricating thin-film nanostructures.
• Measured reflectance spectra to validate model predictions against simulations.

Main Results:

• Color2Struct reduced color difference by 65% and near-infrared reflectance by 48% for sRGB primary colors compared to a baseline tandem-network.
• Experimental validation confirmed the model's predictive accuracy for fabricated nanostructures.
• Demonstrated the framework's ability to provide controllable spectral predictions.

Conclusions:

• Color2Struct offers an efficient and accurate approach for the inverse design of structural colors.
• The framework successfully integrates physical constraints, enhancing controllability of spectral outputs.
• The proposed method shows potential for broad applications beyond nanophotonics.