Electromagnetic Parametric Modeling Using Combined Neural Networks and RLGC-Based Eigenfunctions for Two-Port Microstrip Structures

Abstract

This article proposes a novel electromagnetic (EM) parametric modeling technique using combined neural networks and resistance, inductance, conductance, and capacitance (RLGC)-based eigenfunctions (neuro-EF) for microwave components with two-port microstrip structures. In the proposed technique, neural networks are used to learn the unknown relationship between the parameters of RLGC of the eigenfunction and geometrical parameters. The generation of training data depends on obtaining the correct eigenvalues of different modes. However, for different geometrical parameter samples, there is no uniform correspondence between the calculated eigenvalues and the modes. The incorrect correspondence may cause the two modes to be swapped, defined as the mode-swap issue. We propose a mode-matching method based on eigenvectors to solve this mode-swap issue. After the training data of eigenfunction parameters is obtained, a preliminary training of the neural networks and a two-step refinement training of the neuro-EF model are proposed to develop the overall EM parametric model. By the proposed modeling technique, the trained model can provide a fast and accurate prediction of EM responses for two-port microstrip structures as geometrical parameters change. For the parametric modeling of microwave components with microstrip structures, the proposed technique can obtain better accuracy in larger geometrical variations compared with the existing methods. Two examples of microstrip structures are used to illustrate the proposed technique.