Diffusion probabilistic model based accurate and high-degree-of-freedom metasurface inverse design

Area of Science:

• Electromagnetics and Metamaterials
• Computational Physics
• Machine Learning for Materials Science

Background:

• Traditional metamaterial design relies on inefficient trial-and-error and simulations.
• Existing inverse design algorithms like evolutionary algorithms and topological optimization struggle with multi-objective tasks.
• Generative Adversarial Networks (GANs) show promise for meta-atom inverse design but suffer from training instability and high costs.

Purpose of the Study:

• To develop a novel, stable, and efficient inverse design method for high-degree-of-freedom meta-atoms.
• To address the limitations of existing inverse design techniques, particularly GANs.
• To generate meta-atoms that precisely meet specified S-parameter requirements.

Main Methods:

• A novel inverse design method based on diffusion probability theory is proposed.
• The method learns a Markov process to transform structures into a Gaussian distribution.
• Noise is gradually removed from the Gaussian distribution to generate new meta-atoms.

Main Results:

• The diffusion-based method demonstrates superior model convergence speed compared to GANs.
• The proposed approach achieves higher generation accuracy and quality for meta-atoms.
• It effectively avoids the instability issues inherent in GAN adversarial training.

Conclusions:

• The diffusion probability theory offers a more stable and efficient alternative for metamaterial inverse design.
• This method enables the generation of high-quality meta-atoms meeting complex S-parameter specifications.
• The findings suggest a significant advancement in computational materials science for metamaterial applications.