부산대학교 도서관

This article proposes a novel electromagnetic (EM) parametric modeling approach using model-order reduction (MOR)-based neuro-impedance matrix transfer functions (neuro-IMTFs) for two-port microwave components. The IMTF replaces the impedance matrix response with a zero–pole–gain transfer function, which is directly derived from the finite element method based on MOR. The extracted zeros/poles of the IMTF are located in the frequency space. When we generate training data for parametric modeling, the zeros/poles generated with the change of geometrical parameters can be sorted by the magnitude of the frequency. This fundamentally solves the problem of the difficulty in correctly matching the zeros/poles regenerated every time the geometrical parameters change. We use neural networks to learn the parameterized relationships of sorted zeros/poles and establish the final IMTF model. Finally, the scattering parameter responses of microwave components are obtained from the impedance matrix responses. However, the results of the scattering parameters obtained from the impedance matrix as an intermediate quantity inevitably produce inaccurate peaks. We propose a smoothing algorithm based on pole position and second derivative to solve this problem. The poles of the impedance parameters determine the position to be smoothed, while the second derivative determines the size of the interval to be smoothed. Due to the absence of complex mismatch issues and the resolution of incorrect peaks generated from converting ${Z}$ -parameters to ${S}$ -parameters, the proposed method achieves better model accuracy and more accurate predictions in two-port microwave components involving more considerable geometrical variation ranges for modeling than existing methods. Two examples of two-port microwave components are used to illustrate the proposed approach.