Deep learning-enabled rapid inverse design of high-performance meta-atoms

Conventional metasurface design methods rely on brute-force parameter scanning to optimize structural morphology, while new designs demonstrate limited capability to build upon existing achievements. Although deep learning has significantly advanced this field, current neural networks remain constrained by parameterized regular structures and face challenges in achieving rapid, flexible inverse design for arbitrary-bandwidth configurations. Here, we present a novel end-to-end framework utilizing a vector-quantized variational autoencoder (VQ-VAE) with discrete latent space encoding. The method enables direct mapping from target optical responses to complex, arbitrary metasurface unit geometries, effectively overcoming local optimization. The simulation results demonstrate exceptional performance of the model with 12,000 datasets, achieving a remarkably low mean absolute error of 0.03. Rigorous coupled-wave analysis (RCWA) validation reveals strong correlation coefficients of 0.88 for efficiency and 0.97 for phase response, confirming the model's predictive accuracy. Capable of generating non-conventional meta-atom designs across a wavelength range of 200nm to 1200nm in merely 2 seconds with complete phase coverage from 0 to 2π, this framework offers significant potential for arbitrary band metasurface applications and extensibility for future customization.