Addressing high-performance data sparsity in metasurface inverse design using multi-objective optimization and...

Area of Science:

• Metasurface design
• Deep learning for electromagnetics
• Computational material science

Background:

• Deep learning, especially generative networks, advances meta-atom generation.
• High-performance data scarcity limits current methodologies.
• Metasurface inverse design requires efficient data utilization.

Purpose of the Study:

• To develop a novel approach for meta-atom generation with limited high-performance data.
• To address the challenge of data scarcity in practical metasurface design scenarios.
• To introduce an enhanced diffusion model with attention for improved generation.

Main Methods:

• Synergistic combination of multi-objective optimization algorithms and an enhanced diffusion model (MetaDiffusion-Att) with an attention mechanism.
• Application to the design of dual-polarized, wide-angle incidence, broadband low-emissivity electromagnetic glass.
• Qualitative and quantitative experimental validation.

Main Results:

• Multi-objective optimization captures more high-performance samples with high degrees of freedom compared to generic methods.
• MetaDiffusion-Att outperforms conventional WGAN-GP and conditional VAE on small datasets in generation accuracy and quality.
• The method demonstrates extrapolation capabilities, generating novel structures exceeding dataset performance.

Conclusions:

• The proposed MetaDiffusion-Att framework offers a promising solution for inverse design of metasurfaces.
• It effectively addresses challenges posed by sparse high-performance sample datasets.
• This approach significantly enriches the design space for meta-atom generation.