Fast machine-learning-enabled size reduction of microwave components using response features

Abstract

Achieving compact size has emerged as a key consideration in modern microwave design. While structural miniaturization can be accomplished through judicious circuit architecture selection, precise parameter tuning is equally vital to minimize physical dimensions while meeting stringent performance requirements for electrical characteristics. Due to the intricate nature of compact structures, global optimization is recommended, yet hindered by the excessive expenses associated with system evaluation, typically conducted through electromagnetic (EM) simulation. This challenge is further compounded by the fact that size reduction is a constrained problem entailing expensive constraints. This paper introduces an innovative method for cost-effective explicit miniaturization of microwave components on a global scale. Our approach leverages response feature technology, formulating the optimization problem based on a set of characteristic points derived from EM-analyzed responses, combined with an implicit constraint handling approach. Both elements facilitate handling size reduction by transforming it into an unconstrained problem and regularizing the objective function. The core search engine employs a machine-learning framework with kriging-based surrogates refined using the predicted improvement in the objective function as the infill criterion. Our algorithm is demonstrated using two miniaturized couplers and is shown superior over several benchmark routines, encompassing both conventional (gradient-based) and population-based procedures, alongside a machine learning technique. The primary strengths of the proposed framework lie in its reliability, computational efficiency (with a typical optimization cost ranging from 100 to 150 EM circuit analyses), and straightforward setup.