Designing mm-wave electromagnetic engineered surfaces using generative adversarial networks

Abstract

In this paper, we investigate the capability of generative adversarial networks, including conditional and conditional convolutional generative adversarial networks, in generating electromagnetic engineered surfaces (EES). Generative models such as generative adversarial networks and their conditional variants can be used to generate different categories of designs based on the current dataset. k-means clustering algorithm is used to obtain the desirable categories of EES designs, including an initial two main categories, followed by six and eight subcategories. Conditional and conditional convolutional generative adversarial networks are proposed and trained on designs with different image dimensions conditioned on different sets of categories. The trained conditional convolutional generative adversarial network models have comparable accuracy with conditional generative adversarial network in low-dimensional designs over two categories. Conditional convolutional generative adversarial networks generate more unique designs for six and eight categories for smaller image dimensions (e.g., 9 × 9 designs) and for two main categories over larger designs. Both generative adversarial network structures are suitable for generating a wide variety of low- and high-pass EES designs. The creation of new datasets can benefit from conditional convolutional generative adversarial networks to provide greater variety in designs.