Harnessing the Missing Spectral Correlation for Metasurface Inverse Design

Area of Science:

• Physics
• Computer Science
• Materials Science

Background:

• Deep learning models for electromagnetic metasurfaces often function as "black boxes," lacking physical interpretability.
• Existing intelligent applications in photonics neglect fundamental physical laws, hindering performance and understanding.
• The intrinsic spectral connection between real and imaginary parts, governed by Kramers-Kronig relations, remains largely untapped in deep learning.

Purpose of the Study:

• To reveal and leverage the Kramers-Kronig relations within deep learning frameworks for metasurface inverse design.
• To enhance the feature extraction capabilities of neural networks by incorporating physical spectral connections.
• To provide physically explainable insights into the performance differences of various neural network architectures.

Main Methods:

• Utilized Kramers-Kronig relations to establish an intrinsic correlation between spectral components.
• Implemented a bidirectional information flow in neural network space to mimic these physical relations.
• Benchmarked a bidirectional recurrent neural network against fully-connected networks, unidirectional recurrent neural networks, and attention-based transformers for metasurface inverse design.

Main Results:

• The bidirectional recurrent neural network demonstrated improved accuracy in metasurface inverse design compared to other architectures.
• Analysis of intermediate network products provided physical explanations for performance variations across different network structures.
• The integration of physical laws enhanced the effectiveness of crucial feature extraction.

Conclusions:

• Incorporating physical laws, such as Kramers-Kronig relations, into deep learning offers a more explainable and effective approach for metasurface design.
• Bidirectional information flow in neural networks can successfully mimic physical spectral connections, improving model performance.
• This work provides a pathway for developing data-intensive research endeavors with greater physical insight and transparency.