Abstract
The significance of behavioral models gradually increases in the design and analysis of microwave components. They are mainly used to replace CPU-heavy full-wave electromagnetic (EM) simulations and to expedite EM-driven procedures, especially optimization. Unfortunately, constructing accurate surrogates is a challenging task. In the case of highly nonlinear frequency characteristics of microwave passives, it is normally feasible only when the structures are parametrized by a small number of parameters belonging to narrow ranges. Design utility of such models is limited. Therefore, we developed a novel methodology for computationally efficient and reliable microwave modeling. The presented approach incorporates dimensionality reduction as well as spatial confinement to lower the cost of training data acquisition and to improve the model predictive power. The former is enabled by rapid global sensitivity analysis, which identifies the directions having major influence on the circuit response variability. These directions span the model domain, which is further confined using the pre-screening mechanism focusing on better-quality designs, as well as the spectral analysis of the selected design subset. The surrogate established in the reduced domain still covers the parameter space parts of primary importance, thereby retaining its design applicability. Excellent accuracy of the proposed technique has been validated through extensive benchmarking against several state-of-the-art methods, whereas design readiness has been demonstrated through circuit optimization under various sets of performance requirements. Physical measurements of fabricated circuit prototypes provide auxiliary yet essential validation of the relevance of the proposed modeling technique.