Inverse Design of Metamaterials Using Progressively Trained Variational Autoencoders

Abstract

Inverse design of metamaterials is crucial for advancing nanophotonics, but traditional methods face challenges in pixel-by-pixel optimization and computational efficiency. This work introduces a novel inverse design framework that utilizes a progressively trained variational autoencoder (VAE) in conjunction with data augmentation. Our probabilistic generative model directly maps the desired optical spectrum (reflection, transmission, and absorption) to the optimized metamaterial structure, overcoming the limitation of deterministic methods in handling non-unique mappings. This method enables the rapid and efficient generation of structural design parameters for high-performance nanophotonic components, paving the way for next-generation silicon photonic component design and photonic convolutional neural networks.