Cost-efficient modeling of input characteristics of narrow-band antennas using response features

Modeling of input characteristics of narrow-band antennas is challenging for conventional data-driven approximation techniques such as Gaussian Process regression. The reason is that reflection responses of such structures are highly nonlinear and quickly varying both with respect to frequency and geometry parameters of the antenna. Consequently, a large number of training samples (thus, high CPU cost of data acquisition) is necessary to ensure reasonable predictive power of the model. In this paper, we demonstrate that considerable reduction of the modeling cost can be achieved by utilizing suitably defined characteristic points of the antenna responses (also referred to as response features) because their dependence on the geometry parameters is less nonlinear than for the original responses (S-parameters versus frequency). Numerical comparisons executed for an example dielectric resonator antenna indicate that accurate feature-based surrogate models can be constructed at a fraction of the computational cost required by conventional modeling techniques.