Surrogate Modeling of High-Frequency Structures Using Nested Kriging and Improved Sampling Strategy

Abstract

Fast surrogate models are indispensable in reducing the computational cost of electromagnetic (EM)-driven design processes of high-frequency structures. The most popular class of surrogates, data-driven (or approximation) models suffer from the curse of dimensionality, which limits both the number of parameters and their ranges that can be effectively handled. A recently proposed nested kriging modeling framework allows us to work around these limitations by restricting the surrogate model domain to the vicinity of a set of reference designs pre-optimized with respect to the selected performance figures. This significantly improves the model reliability while reducing the number of training data samples necessary to set up the surrogate. In this paper, an improved design of experiments (sampling scheme) is proposed, which further reduces the modeling error as demonstrated using two high-frequency structures: a dual-band uniplanar dipole antenna, and a miniaturized impedance matching transformer.