Low-cost and precise automated re-design of antenna structures using interleaved geometry scaling and gradient-based optimization
Anna Pietrenko-Dabrowska,Slawomir Koziel
DOI: https://doi.org/10.1016/j.knosys.2023.111296
IF: 8.139
2024-01-02
Knowledge-Based Systems
Abstract:Design of contemporary antennas is an intricate endeavor involving multiple stages, among others, tuning of geometry parameters. In particular, re-designing antennas to different operating frequencies, makes parametric optimization imperative to ensure the best achievable system performance. If the center frequency at the current design is distant from the target one, local tuning methods generally fail, whereas global algorithms (e.g., nature-inspired procedures) incur prohibitive computational expenses, especially when antenna evaluation is performed using full-wave electromagnetic (EM) analysis. In this paper, a novel technique involving automated decision-making has been developed, whose main objective is low-cost and precise re-design of antenna structures over wide ranges of operating frequencies. The employed methodology involves knowledge-based simultaneous scaling of antenna dimensions and gradient-based performance improvements. The two stages are automatically interleaved, and embedded into an iterative optimization procedure . The problem-specific knowledge allows for carrying out the scaling phase, in which fast relocation of the center frequency of the antenna is performed, based on a single EM analysis of the structure. The gradient-based tuning phase enhances the design quality with regard to the assumed objectives. The process defaults to local optimization after the antenna center frequency becomes sufficiently close to the target. The main novelty of the proposed algorithm consists in development of an automated knowledge-based framework of quasi-global search capabilities linking brute-force scaling and design refinement. Our technique has been demonstrated with the use of three microstrip antennas, optimized for best matching and maximum in-band gain. The main findings are that for all structures, satisfactory designs have been identified despite poor starting points, with operating frequencies being away from the assumed targets. At the same time, the computational cost is comparable to conventional local search. The proposed approach is versatile, simple to implement and easy to handle, in particular, its control parameters do not require tailoring to a specific antenna structure at hand.
computer science, artificial intelligence