High-Asymmetry Metasurface: A New Solution for Terahertz Resonance via Active Learning-Augmented Diffusion Model

Area of Science:

• Metamaterials Science
• Artificial Intelligence in Physics
• Terahertz Technology

Background:

• High-figure-of-merit (high-FoM) Terahertz (THz) metamaterials are crucial for advanced sensors, detectors, and imagers.
• Conventional metamaterial design often relies on symmetric structures, limiting exploration of high-asymmetry designs due to inefficient trial-and-error methods.
• Current deep learning approaches for metamaterial design are hindered by substantial data requirements.

Purpose of the Study:

• To develop a novel generative model for designing high-asymmetry Terahertz metamaterials with high-FoM resonance.
• To overcome the data limitations of traditional deep learning methods in metamaterial design.
• To enhance the performance of THz metadevices through innovative structural design.

Main Methods:

• A prior knowledge-guided generative model, specifically an advanced diffusion model, was employed.
• The model learned features from a small dataset of classical high-FoM THz resonant structures.
• A physics-constrained active learning mechanism was integrated to select and augment the training dataset with promising generated structures.

Main Results:

• The generative model successfully designed novel high-asymmetry metamaterial structures.
• Experimental validation showed superior resonance performance of the generated high-asymmetry metamaterials compared to classical designs, with key metrics improving by over 30%.
• The design process achieved these results using a remarkably small initial training dataset of only 68 classical structures.

Conclusions:

• The proposed method offers an effective and efficient solution for designing high-FoM asymmetric metamaterials.
• This approach significantly reduces the data dependency for deep learning-based metamaterial design.
• The findings pave the way for broader applications of high-sensitivity THz metadevices.