Abstract
Rigorous numerical optimization is ubiquitous in modern antenna design. In many cases, performing local (e.g., gradient-based) parameter tuning is sufficient. However, a global optimization is necessary in many practical scenarios, which incurs tremendous computational expenses and is often unmanageable, mainly when carried out using electromagnetic (EM) analysis. It is possible to mitigate this issue using surrogate modeling techniques. Still, building quality metamodels is hindered by the curse of dimensionality and the necessity of setting up the models across extended ranges of geometry parameters. This paper presents a cost-effective technique for globalized antenna optimization. Our approach carries out the search process in the space of antenna operating parameters (e.g., center frequencies), using simplex-based regression predictors and variable-resolution EM simulations. The stage of global optimization, conducted using low-fidelity EM analysis, is complemented by rapid gradient-based tuning performed using high-resolution models. The latter is accelerated by performing the antenna sensitivity only along certain (principal) directions, affecting response variability to the greatest extent. Comprehensive validation using four microstrip antennas shows the excellent performance of the presented method and its superior computational efficiency over several benchmark methods (less than eighty high-fidelity EM simulations are required to render optimal design on average).