Inverse design of meta-surfaces based on generative adversarial networks

Metasurfaces, as artificial structures capable of precise manipulation of optical wavefronts, demonstrate immense potential in the fields of optoelectronics and wireless communications. However, traditional design methods for metasurfaces often rely on intuition and trial-and-error, which are not only inefficient but also struggle to meet complex design requirements. The rise of deep learning technologies in recent years has offered a new perspective for addressing this issue. This study designs and improves a novel deep convolutional network architecture that integrates Wasserstein GANs and attention mechanisms for the efficient inverse design of metasurface structures. After testing on internal and external test sets and random spectra, the attention-based WGAN network model achieved an image accuracy of 0.98 and a spectral accuracy of 0.97, proving its efficiency and accuracy in designing elliptical metasurface structures.