Machine learning for microwave optimization using simplex surrogates, dual-resolution computational models and local tuning with sparse sensitivity updates

Abstract

Abstract Numerical optimization procedures are now an integral part of the microwave design process. Ensuring reliability requires conducting parameter tuning at the electromagnetic (EM) analysis level. This, however, entails considerable computational costs. Additionally, global optimization is often necessary (e.g., multimodal problems, large-scale operating frequency re-design, design of metasurfaces), which is incomparably more expensive when using conventional techniques. In this work, we introduce a novel approach to the fast globalized optimization of microwave structures. Our methodology is founded on processing the operating parameters of the circuit rather than its complete frequency characteristics, and the utilization of simplex-based regressors. Both permit regularizing the objective function, which facilitates and speeds up the identification of the optimum design. Further acceleration is enabled by employing dual-fidelity EM simulations and restricted sensitivity updates at the final parameter tuning stage. The introduced algorithm has been comprehensively demonstrated using several microstrip components and proved to be superior over several benchmark approaches. Apart from reliability, its attractive features include remarkable computational efficiency (the average optimization cost corresponding to fewer than fifty EM simulations of the circuit), as well as simple implementation and handling with a small number of control parameters that do not have to be tuned to a specific problem at hand.