Guided by Uncertainty: Adaptive Frequency Sampling Using Gaussian Processes

Abstract

Adaptive frequency sampling (AFS) aims to reduce the number of computationally expensive electromagnetic (EM) simulations required for the accurate characterization of high-frequency components. This paper introduces a novel uncertainty-guided AFS scheme that employs a rational mixture kernel within a Gaussian Process (GP) model to capture frequency-dependent correlation structures in microwave Sparameters. By leveraging the probabilistic nature of GPs, an acquisition strategy is introduced to adaptively sample uncertain regions exhibiting high local variation. Integrated within the kernel-aided rational macromodeling (KARMA) framework, the proposed approach enables the construction of compact rational state-space models with enhanced accuracy and reduced simulation cost.