Koziel Slawomir, Pietrenko-Dabrowska Anna
Engineering Optimization & Modeling Center, Reykjavik University, 101, Reykjavik, Iceland.
Faculty of Electronics, Telecommunications and Informatics, Gdansk University of Technology, Gdansk, 80-233, Poland.
Sci Rep. 2024 Dec 28;14(1):31265. doi: 10.1038/s41598-024-82701-3.
Recent years have witnessed a tremendous popularity growth of optimization methods in high-frequency electronics, including microwave design. With the increasing complexity of passive microwave components, meticulous tuning of their geometry parameters has become imperative to fulfill demands imposed by the diverse application areas. More and more often, achieving the best possible performance requires global optimization. Unfortunately, global search is an intricate undertaking. To begin with, reliable assessment of microwave components involves electromagnetic (EM) analysis entailing significant CPU expenses. On the other hand, the most widely used nature-inspired algorithms require large numbers of system simulations to yield a satisfactory design. The associated costs are impractically high if not prohibitive. The use of available mitigation methods, primarily surrogate-based approaches, is impeded by dimensionality-related problems and the complexity in microwave circuit characteristics. This research introduces a procedure for expedited globalized parameter adjustment of microwave passives. The search process is embedded in a surrogate-assisted machine learning framework that operates in a dimensionality-restricted domain, spanned by the parameter space directions being of importance in terms of their effects on the circuit characteristic variability. These directions are established using a fast global sensitivity analysis procedure developed for this purpose. Domain confinement reduces the cost of surrogate model establishment and improves its predictive power. The global optimization phase is complemented by local tuning. Verification experiments demonstrate the remarkable efficacy of the presented approach and its advantages over the benchmark methods that include machine learning in full-dimensionality space and population-based metaheuristics.
近年来，优化方法在包括微波设计在内的高频电子学中广受欢迎。随着无源微波元件复杂性的增加，对其几何参数进行精细调整对于满足不同应用领域的要求变得至关重要。越来越多地，要实现最佳性能需要全局优化。不幸的是，全局搜索是一项复杂的任务。首先，对微波元件进行可靠评估涉及电磁（EM）分析，这需要大量的CPU开销。另一方面，最广泛使用的自然启发算法需要大量的系统仿真才能得到令人满意的设计。如果不是过高的话，相关成本也高得令人望而却步。可用的缓解方法，主要是基于代理的方法，受到与维度相关的问题以及微波电路特性复杂性的阻碍。本研究介绍了一种用于微波无源器件快速全局参数调整的程序。搜索过程嵌入在一个代理辅助的机器学习框架中，该框架在一个维度受限的域中运行，该域由对电路特性变化有重要影响的参数空间方向所跨越。这些方向是使用为此目的开发的快速全局灵敏度分析程序确定的。域限制降低了代理模型建立的成本并提高了其预测能力。全局优化阶段辅以局部调整。验证实验证明了所提出方法的显著有效性及其相对于包括全维空间机器学习和基于群体的元启发式算法在内的基准方法的优势。
