Kernel-Based Machine Learning Surrogates for RF Circulator Design

Machine learning has attracted significant interest in the microwave community in recent years. It has the potential to greatly enhance design flows, but the choice of the algorithm and, more importantly, of the deployment strategy are nontrivial and may seem intimidating with respect to traditional approaches. This paper is intended as a proof of concept. It demonstrates a simple, effective, and elegant way to build and use feature-based surrogate models in order to improve the design process of RF circulators. The models are able to predict key metrics of the frequency response of the device, depending on its geometry. These models are then used to select configurations that meet the desired specifications.