Globalized parameter tuning of microwave passives by dimensionality-reduced surrogates and multi-fidelity simulations

Parameter tuning is an essential but demanding aspect of microwave component design, particularly when global optimization is required. The process becomes especially demanding due to the extensive electromagnetic (EM) simulations involved, which-when using popular nature-inspired methods-can lead to unmanageable computational costs. Traditional mitigation approaches, such as surrogate-based methods, often struggle with the curse of dimensionality and the highly nonlinear responses of microwave circuits. This study introduces an alternative approach for rapid global optimization of microwave passive components using artificial intelligence (AI) techniques, specifically, machine learning (ML). The core elements of our methodology include reduction of the problem dimensionality using a rapid global sensitivity analysis, multi-fidelity EM simulations, and a two-stage search process. During the global optimization stage, surrogate-assisted ML is confined to a reduced-dimensionality region, leading to significant computational savings and enhanced predictive accuracy of the surrogate models. Additional speedup is achieved by performing the search using low-fidelity EM models. The final local refinement stage employs high-resolution models and is executed within the design space of full dimensionality, ensuring the quality of the final design. Our procedure was comprehensively validated using four microstrip circuits and has demonstrated superiority over state-of-the-art benchmark methods. The average optimization cost is equivalent to only about ninety EM simulations. Further, the quality of the resulting designs remains competitive with those rendered using the benchmark methods.