Inverse Design of a Dual-Band Reflective Polarizing Surface Using Generative Machine Learning

Reflective linear-to-circular polarization (LP-to-CP) converters are desired for satellite communication applications to reduce the complexity of the antenna system and the number of the required apertures in single-feed-per-beam configurations. Moreover, wideband operation and angular stability are important polarizer attributes in these applications. Here, for the first time, we use a machine-learning technique employing a generative adversarial network (GAN) to systematically and efficiently explore and exploit the solution space of single-layer structures for this problem. By using a GAN to optimize the shape of the polarizing element, superior performance in terms of overlapping axial ratio bandwidth for normal and oblique incidence of 30 degrees compared to the previously studied designs is achieved.

Automated summary

This paper utilizes a machine-learning technique called generative adversarial network (GAN) to explore and exploit the solution space of reflective linear-to-circular polarization (LP-to-CP) converters. By optimizing the shape of the polarizing element, the proposed designs achieve superior performance in terms of overlapping axial ratio bandwidth for normal and oblique incidence of 30 degrees compared to previously studied designs.