Reliable Computationally Efficient Behavioral Modeling of Microwave Passives Using Deep Learning Surrogates in Confined Domains

The importance of surrogate modeling techniques has been steadily growing over the recent years in high-frequency electronics, including microwave engineering. Fast metamodels are employed to speed up design processes, especially those conducted at the level of full-wave electromagnetic (EM) simulations. The surrogates enable massive system evaluations at nearly EM accuracy and negligible costs, which is invaluable in parameter tuning, multiobjective optimization, or uncertainty quantification. Nevertheless, modeling of electrical characteristics of microwave components is impeded by nonlinearity of their electrical characteristics, the need for covering broad parameter ranges, as well as dimensionality issues. Recently, a two-stage modeling approach has been proposed, which addresses some of these issues by constraining the surrogate model domain to only include high-quality designs, thereby reducing the cardinality of the dataset required to establish an accurate metamodel. In this article, a novel technique is proposed, which combines the two-stage modeling concept with multihead deep regression network (MHDRN) surrogates customized to handle responses of microwave passives over wide ranges of operating frequencies and geometry parameters. Using three microstrip circuits, a superior performance of the proposed modeling framework is demonstrated with respect to multiple state-of-the-art benchmark methods. In particular, the relative rms error is shown to reach the level of less than 3% for the datasets consisting of just a few hundred samples.