A Generative Adversarial Network-Based Approach for Reflective-Metasurface Unit-Cell Synthesis in mmWave Bands

In recent years, data-driven design approaches based on generative machine learning (ML) have been applied in the field of metasurface design. Even though generative machine learning has demonstrated remarkable capabilities in metasurface synthesis, training procedures to obtain an appropriate model remain a challenge. In this work, we present the deployment of conditional generative adversarial networks (cGAN) to design reflective standard unit cells for the use of 5G millimeter wave (mm Wave) passive metasurface reflectors. A total of 48 different configurations of cGAN were trained over a range of hyperparameters selected under proper guidelines used in the literature. The design performance of each cGAN configuration is measured by a mean absolute error (MAE) between a simulated reflection phase of a generated unit cell and a target reflection phase. The test results show that cGAN s trained by updating the discriminator one step for each single generator update tend to offer lower MAEs compared to cGANs trained by updating the discriminator three iterations for one generator update. The optimal cGAN among 48 hyperparameter configurations can generate unit cells with reflection phase resembling the desired target one.