Cost-Efficient Globalized Parameter Optimization of Microwave Components Through Response-Feature Surrogates and Nature-Inspired Metaheuristics

Design of contemporary microwave devices predominantly utilizes computational models, including both circuit simulators and full-wave electromagnetic (EM) evaluation. The latter constitutes the sole generic way of rendering accurate assessment of the system outputs that considers phenomena such as cross-coupling or radiation and dielectric losses. Consequently, for reliability reasons, the final tuning of microwave device parameters is commonly performed utilizing EM simulation software. As EM analysis is computationally heavy, parametric optimization entails significant costs, also for local algorithms. The expenses generated by global search procedures are incomparably higher, and often prohibitive. Still, global optimization is more and more often necessary, for example, when re-designing a structure over extended ranges of operating conditions, if more than a single local optimum exists, or simply due to the absence of quality initial design. A possible workaround is surrogate-assisted optimization, yet a construction of accurate replacement models is a challenge by itself. This paper offers an innovative approach to a rapid globalized optimization of passive microwave components. It combines a machine learning procedure, specifically, an iterative construction and refinement of fast surrogates (with infill criterion being a minimization of the predictor-yielded objective improvement) with a response feature technology, where the metamodel targets suitably appointed characteristic points of the circuit outputs. These so-called response features are in a nearly linear relationship with the geometry parameters, which facilitates the search process and reduces the expenditures associated with surrogate model construction. Identification of the infill points is executed using a particle swarm optimization algorithm. Numerical experiments carried out using two microstrip circuits demonstrate the capability for a global search of the proposed algorithm, and its superior performance over direct nature-inspired-based optimization and surrogate-assisted search at the level of complete circuit characteristics. The original contributions of this work can be summarized as follows: (i) the development and implementation of the machine learning procedure that operates at the level of response features, (ii) the development of parameter space pre-screening stage employed to narrow down the region to be explored in the search process, (iii) demonstration of superiority of the proposed framework (including its remarkable computational efficiency) over a range of benchmark methods, both direct and surrogate-assisted ones.