Comparison of Circuit Models for ML-Assisted Microwave Circuit Design

Machine-learning (ML) assisted microwave circuit design is an interesting complement to traditional topology-based design since it opens up previously unexplored design spaces that in some cases may offer better performance, or similar performance with a different form factor. A key part is the circuit model, i.e., the set of discrete building blocks used to create circuits. In the work published so far circuit models encompassed a single element type in the form of metal pixels. In this paper we propose a circuit model with additional elements that facilitates diagonal connections and provides higher robustness to variations in the manufacturing process. A comparison with the pixel model shows that the new model results in more accurate ML-models for S-parameter prediction with a 9.5% reduction in root mean-square error (RMSE) on the testset, which translates to more accurate results for circuit synthetization. In addition, we demonstrate that circuits built with the new model has a higher tolerance to manufacturing imperfections, with 33% smaller RMSE penalty with respect to the original S-parameters when adding a width perturbation of 50 $\mu$m to diagonal connections, and 50/40% smaller RMSE penalty when shrinking/expanding the size of elements forming diagonal connections with 2.5% . We also use both the pixel model and the newly proposed model to design low-pass filters with competitive performance.