Inverse Design of Bifunctional Metasurfaces Using Improved Generative Adversarial Networks

Abstract

In this letter, a novel inverse design method for metasurfaces (MSs) based on a Wasserstein generative adversarial network with a gradient penalty (WGAN-GP) is presented. Compared with other GAN variants, the proposed WGAN-GP significantly improves the stability and robustness of the training process using the Wasserstein distance and gradient penalty to ensure a smoother optimization landscape. Furthermore, Gramian angular difference fields (GADFs) are introduced to transform electromagnetic (EM) responses into 2-D images. GADFs are characterized by capturing repetitive patterns and structures in a 1D sequence, making them particularly suitable for processing periodic phase data. Therefore, meta-atom patterns and their corresponding EM responses form 2-D input-output pairs, allowing the WGAN-GP to inversely design MSs from an image recognition perspective. As a proof-of-concept example, we experimentally demonstrate a bifunctional MS that integrates second-order orbital angular momentum (OAM) and holographic imaging under dual-linearly polarized excitation. The measured results closely align with the simulated results, thereby validating the feasibility of our inverse design strategy.