Abstract
The development of contemporary electronic components, particularly antennas, places significant emphasis on miniaturization. This trend is driven by the emergence of technologies such as mobile communications, the internet of things, radio-frequency identification, and implantable devices. The need for small size is accompanied by heightened demands on electrical and field properties, posing a considerable challenge for antenna design. Shrinking physical dimensions can compromise performance, making miniaturization-oriented parametric optimization a complex and heavily constrained task. Additionally, the task is multimodal due to typical parameter redundancy resulting from various topological modifications in compact antennas. Identifying truly minimum-size designs requires a global search approach, as the popular nature-inspired algorithms face challenges related to computational efficiency and the need for reliable full-wave electromagnetic (EM) simulation to evaluate device's characteristics. This study introduces an innovative machine learning procedure for cost-effective global optimization-based miniaturization of antennas. Our technique includes parameter space pre-screening and the iterative refinement of kriging surrogate models using the predicted merit function minimization as an infill criterion. Concurrently, the design task incorporates design constraints implicitly by means of penalty functions. The combination of these mechanisms demonstrates superiority over conventional techniques, including gradient search and electromagnetic-driven nature-inspired optimization. Numerical experiments conducted on four broadband antennas indicate that the proposed framework consistently yields competitive miniaturization rates across multiple algorithm runs at low costs, compared to the benchmark.