Inverse design of structural color: finding multiple solutions via conditional generative adversarial networks

Area of Science:

• Nanophotonics
• Machine Learning
• Computational Physics

Background:

• The
• one-to-many
• problem in machine learning-aided inverse nanophotonics design leads to discarding potential solutions.
• Existing methods like tandem networks and mixture density models do not fully resolve solution degeneracy.

Purpose of the Study:

• To address the
• one-to-many
• problem by developing a novel conditional generative adversarial network (cGAN) approach.
• To enable the generation of multiple, distinct solution groups for inverse design problems.

Main Methods:

• Employed a conditional generative adversarial network (cGAN) for inverse nanophotonics design.
• Utilized a transmissive Fabry-Pérot-cavity-based color filter for demonstrating the model.
• Generated multiple solution groups for each target color in the inverse design process.

Main Results:

• The cGAN model successfully generated an average of 3.58 solution groups per color for the inverse design of color filters.
• Achieved a record high accuracy with an average color difference (ΔE) of 0.44.
• Experimental verification confirmed the cGAN's capability in inversely designing RGB color filters.

Conclusions:

• The cGAN approach effectively solves the
• one-to-many
• problem in inverse design by providing multiple viable solutions.
• This methodology enhances design flexibility and manufacturing feasibility in nanophotonics.
• The cGAN framework is applicable to diverse scientific domains requiring multi-solution identification within large parameter spaces.