Deep residual and variational autoencoding networks with fine-grained parameter optimization: a comprehensive and efficient framework for random polygonal metasurface design

This study presents a deep learning framework for the forward and inverse design of metasurface unit structures, integrating a residual network (ResNet)-based forward predictor, a conditional variational autoencoder (CVAE)-based inverse generator, and a hierarchical hyperparameter optimization strategy. The forward model, enhanced with residual connections and dual-convolution modules, achieves over 96% prediction accuracy. The inverse model, using a CVAE architecture with a dual-convolution decoder, generates high-dimensional structural encodings from target spectra. The framework is validated through two metasurface design examples: a phase-modulation metasurface for the visible spectrum and a multi-wavelength achromatic metalens in the mid-infrared range, demonstrating strong adaptability and efficiency across spectral ranges and functional tasks.