A Generative Adversarial Network Approach to Reflectarray Pattern Synthesis

In this paper, we improve the structure of a generative adversarial network (GAN) based on three fundamental characteristics of the reflectarray for synthesizing better reflectarray patterns. For meeting various challenging requirements of beamforming, we propose a new model, which uses different tactics for improvements in the generator part of a GAN model. First, for accommodating the error-sensitive characteristics of the reflectarray antenna, a fully connected network is used to increase model precision. Second, based on the effects of interference occurring among all elements in a reflectarray, a self-attention mechanism is introduced to better handle the relationship among pixels. Third, with the needs of representing the associated values of elements based on the actual physical distance between the feed horn and elements, interpolation is used to enlarge the image, and then additional convolutional layers are provided to adjust the details. The results show acceptable radiation patterns produced by the GAN model.