Adversarial-Network Regularized Inverse Design of Frequency-Selective Surface With Frequency-Temporal Deep Learning

In contrast to designing electromagnetic (EM) structures by tuning computationally expensive forward models, recent advancements in data-driven inverse-design approaches have potentials to transform the designs to more effective and intelligent ways. However, the fundamental challenges of the nonuniqueness in mapping from the demand space to geometry space and the difficulty in incorporation of fabrication constraints severely hinder the practical applications of these data-driven methods. To alleviate both challenges, we propose an adversarial-network regularized inverse-design scheme with frequency-temporal deep learning method (AR-FTDL). A key feature of the proposed AR-FTDL is that it initially generates data distributions from the demand functional characteristics regularized by a generative adversarial network (GAN). Furthermore, the inversion network utilizes both frequency and time prior information of data space, which is followed by a pseudo-twin network to constrain geometry space satisfying fabrication requirements. The proposed method is numerically and experimentally verified by inversely designing passive/active EM absorbers with absorption–transmission–absorption functions. We believe that the proposed data-driven scheme can extend and improve the capabilities of the traditional intuition-driven or intense computation-driven design method.