An Efficient Sensitivity-Driven Stepwise Method Incorporating Transfer Learning for Wide-Range Parametric Modeling of Microwave Components

Abstract

This letter proposes an efficient sensitivity-driven stepwise modeling method for microwave components with a wide range of geometrical parameter variations. In the proposed method, the Pearson correlation coefficient is explored to solve the sensitivity analysis difficulty in accurately classifying the geometrical parameters into high-sensitivity parameters and low-sensitivity parameters. The relationship between high-sensitivity parameters and circuit responses is learned by the neural network as the first modeling step. Then, the effect of low-sensitivity parameters on the circuit response is restored in the second modeling step through transfer learning (TL), which leverages the knowledge gained from high-sensitivity parameters. Using the proposed sensitivity-driven stepwise modeling method allows us to achieve a much faster training convergence speed through effective knowledge transfer and reuse, consequently achieving similar accuracy in a shorter training time compared with existing methods using the same data. Two microwave modeling examples are used to illustrate the proposed method.