On reliable cost-efficient data-driven modelling of microwave passives in reduced-dimensionality domains with co-kriging

Data-driven surrogate modeling has seen a rise in attention in microwave engineering over the recent years. It has been primarily employed to expedite electromagnetic (EM)-based design procedures, e.g., parametric optimization or robust design. However, behavioral modeling of microwave components becomes problematic due to their nonlinear characteristics, often large numbers of geometry parameters, as well as broad ranges thereof. This work proposes an unconventional approach to resource-efficient and accurate modeling of microwave passive circuits. Our methodology is founded on reducing the design space dimensionality carried out via fast global sensitivity analysis, and also the incorporation of variable-resolution simulation data. The surrogate domain extends alongside the most significant directions reflecting the highest circuit response variability. The surrogate model is rendered using a mixture of densely sampled low- and sparsely allocated high-resolution EM simulation training points, which are blended together using co-kriging. The aforementioned mechanisms ensure remarkable computing performance of the modeling process and greatly improve the surrogate’s predictive power as compared to surrogates established in a full-dimensionality domain. Comparative studies carried out using three circuits and a number of recent benchmark methods corroborate the competitive performance of the introduced methodology, in particular, its ability to yield design-ready behavioral models featuring relative RMS error of a few percent at the cost corresponding to less than a few hundred of high-fidelity training points.