Diffusion model-based inverse design of photonic crystals for customized refraction

Area of Science:

• Photonics and Materials Science
• Artificial Intelligence in Engineering

Background:

• Photonic crystals (PhCs) are crucial for integrated photonic systems.
• Traditional PhC design methods lack efficiency and flexibility.
• Existing deep learning models (GANs, VAEs) face training instability and noise issues.

Purpose of the Study:

• To introduce a novel generative design framework for high-precision, customized PhC refraction structures.
• To address limitations in current deep learning approaches for PhC inverse design.

Main Methods:

• Developed a diffusion model-based generative framework for PhC inverse design.
• Constructed a dataset of PhC equifrequency contours (64x64 resolution) including frequency, angles, and patterns.
• Integrated the diffusion model with the U-Net architecture for structure generation.

Main Results:

• Successfully generated customized PhC structures with high precision.
• Achieved accurate prediction of cell patterns for incident angles (0°-80°) and refraction angles (-80° to 80°).
• 85% of 1000 tested structures showed refracted angle errors below 0.1 (L2-norm), demonstrating high stability and precision.

Conclusions:

• The diffusion model-based framework offers a promising solution for automated inverse design of photonic devices.
• The approach provides high stability, precision, and adaptability for multi-scale photonic applications.
• Expanding the dataset can further enhance the solution space for diverse PhC designs.