Wide-Range Parametric Modeling of Microwave Components

Wide-range parametric modeling of microwave components can significantly expand the applicability of models. However, traditional neural network modeling methods face challenges, such as high nonlinearity and low training efficiency. This article analyzes the explorations and limitations of existing methods when expanding the parameter range and highlights two innovative methods specifically designed for wide-range parameters. The first method is a parallel decomposition approach based on second-order derivative information. It decomposes the parameter domain into subregions, constructs submodels using parallel training and correction, and ultimately integrates them into an overall model. The second method is a sensitivity-driven stepwise approach that incorporates transfer learning (TL). It classifies geometric parameters using the Pearson correlation coefficient and employs TL technology for step-by-step training to achieve a final accurate model. The article also reviews application examples corresponding to both methods, demonstrating their effectiveness. Each method has its own advantages. Future research can explore hybrid strategies of the two methods to leverage their respective strengths and further advance microwave modeling technology.