Deep residual and variational autoencoding networks with fine-grained parameter optimization: a comprehensive and...

Area of Science:

• Photonics and Metamaterials
• Artificial Intelligence in Optics
• Computational Electromagnetics

Background:

• Metasurfaces offer unprecedented control over light but their design is complex.
• Traditional design methods are often time-consuming and limited in scope.
• Deep learning presents a promising avenue for accelerating metasurface design.

Purpose of the Study:

• To develop a unified deep learning framework for both forward and inverse design of metasurface unit structures.
• To enhance prediction accuracy and generation efficiency for diverse optical functionalities.
• To validate the framework's adaptability across different spectral ranges and design tasks.

Main Methods:

• A residual network (ResNet)-based forward predictor with residual connections and dual-convolution modules.
• A conditional variational autoencoder (CVAE)-based inverse generator featuring a dual-convolution decoder.
• A hierarchical hyperparameter optimization strategy for model tuning.

Main Results:

• The forward model achieved over 96% prediction accuracy for metasurface properties.
• The inverse model successfully generated high-dimensional structural encodings from target spectra.
• The framework demonstrated strong adaptability and efficiency in designing a phase-modulation metasurface and an achromatic metalens.

Conclusions:

• The proposed deep learning framework provides an efficient and accurate solution for metasurface unit structure design.
• The integrated approach of forward prediction and inverse generation accelerates the discovery of novel metasurface functionalities.
• The framework shows significant potential for advancing metasurface applications across visible and mid-infrared spectra.