Globalized Machine-Learning-Enabled Antenna Miniaturization Involving Simplex-Metamodels

This work proposes an innovative algorithmic framework for rapidly reducing antenna size through constrained global optimization. The three main stages of the search process are the size-and feature-based pre-screening, a machine learning optimization involving simplex-anchored regression models, and final gradient-based tuning. The first stage allows the determination of the promising region of the decision variable space. The simplex-based predictors constructed and processed therein allow us to rapidly approach a good-quality design, further refined in the subsequent stage. Extensive numerical experiments demonstrate superiority of our methodology over a broad range of benchmark methods regarding the attainable miniaturization and its remarkable efficiency. The average optimization cost corresponds to just 170EM antenna analyses. Meanwhile, the reliability surpasses that of most techniques included in the comparative studies. Potential limitations of the presented approach is in scaling the method for higher-dimensional problems, where the computational cost (especially of the pre-screening stage) will likely be inflated while searching for a set of good-quality initial points employed to jump-start the machine learning process. These issues will be addressed in future work.