Simplex-anchored regressors for fast global optimization-oriented miniaturization of microwave circuits

Abstract

The small physical size of high-frequency components is an essential prerequisite in an increasing number of applications. While size reduction of microwave circuits can be achieved with a suitable choice of architecture, ensuring the best available compromise between electrical parameters and size necessitates meticulous adjustment of all system dimensions. Due to the complexity of electrically small structures, parameter adjustment should be executed globally. Nevertheless, it is an expensive endeavor because it entails a large number of system evaluations executed through electromagnetic (EM) simulation. This article introduces an innovative approach to expedited miniaturization of microwave passives, in which the primary goal is miniaturization, while the conditions imposed on performance metrics are treated as constraints. Our framework operates in three stages, including parameter space pre-screening, machine-learning (ML)-driven globalized search, and gradient-based fine-tuning. The first stage involves random sampling aimed at identifying the most promising parameter space region. The ML process launched therein employs simplex-based regression models, the use of which contributes to a competitive computational efficiency. On the other hand, the final tuning improves design quality at a low cost, as it is realized as a local optimization process. The performance of the suggested algorithm is illustrated using two planar devices and demonstrated to be superior to several benchmark methods. The major advantages of our technique include low running costs, equivalent to only a few dozen full-wave analyses, and straightforward implementation.