A Microwave Filter Yield Optimization Method Based on Off-Line Surrogate Model-Assisted Evolutionary Algorithm

Most existing microwave filter yield optimization methods target a small number of sensitive design variables (e.g., around 5). However, for many real-world cases, more than ten sensitive design variables need to be considered. Due to the complexity, yield optimization quality and efficiency become challenges. Hence, a new method, called yield optimization for filters based on the surrogate model-assisted evolutionary algorithm (YSMA), is proposed. The fundamental idea of YSMA is to construct a single high-accuracy surrogate model offline, which fully replaces electromagnetic (EM) simulations in the entire yield optimization process. Global optimization is then enabled to find designs with substantial yield improvement efficiently using the surrogate model. To reduce the number of necessary samples (i.e., EM simulations) while obtaining the required prediction accuracy, a customized machine learning technique is proposed. The performance of YSMA is demonstrated by two real-world examples with 11 and 14 design variables, respectively. Experimental results show the advantages of YSMA compared to the current dominant sequential online surrogate model-based local optimization methods.