Parametric Modeling of Microwave Structure with Customization Responses by Combining RBF Neural Network and Pole-Residue-Based Transfer Functions

Abstract

This paper proposed a parametric modeling technique for the microwave structures with a customization magnitude response by combining the RBF neural network and pole-residue-based transfer functions. The Latin hypercube sampling method is used for sampling given physical ranges and obtaining the EM behaviors of the microwave structures. A pole sorting process and a modified pole-residues splitting process are proposed to solve the pole sequence chaos and order-changing problems which occur in the modeling process. The pole-residues parameters after the above preprocessing steps are used as the inputs of the RBF neural network and the physical parameters are used as the outputs of RBF network. Then, the known magnitude response of the microwave structure are used as the prior knowledge to guide obtaining the goal pole-residues values corresponding to the giving magnitude response specification. After the training process of the RBF model, the goal pole-residues are input into the trained RBF network and the goal physical parameters corresponding to the customization responses is obtained. Finally, this model technique is illustrated by the two examples of microwave structures.