Neural network based techniques for microwave passive components modeling and design

This thesis presents the neural network based techniques for microwave passive components modeling and design. First, a modeling method called combined approach of transfer function and neural network (TF-NN) is described. Instead of using equivalent circuit, it uses transfer function to provide prior knowledge to the neural network such that the complexity of the neural network input-output mapping for the target problem is simplified. This reduction in mapping complexity helps to develop a neural network model with less training data while keeping the electromagnetic (EM) level accuracy. This TF-NN model is much faster than detailed EM simulation and is more flexible for general passive component modeling than equivalent circuit based knowledge model. Passive microwave components such as via and square inductor models are developed using the TF-NN method. Secondly, another new approach, which is aimed at designing one of the basic microwave components, microstrip power divider, is also proposed. The EM structure of the power divider is first segmented in order to simplify the structure. Detailed EM simulation time is reduced through the model segmentation. The neural networks are used to learn the EM behaviour of the power divider's sub structures. The trained neural sub models are then combined to form the overall power divider structure which can be used during model optimization and provide fast evaluation of the power divider's structure. The optimization time is significantly reduced by using the proposed design approach.