Parametric modeling of electromagnetic (EM) behaviors has become important for EM design optimizations of microwave components. The EM based design, such as design optimization, what if analysis and yield-driven design, can be time consuming because it usually requires repetitive EM simulations with varying values of geometrical parameters as design variables. Parametric models can be developed from the information of EM responses as functions of geometrical parameters. The developed parametric models allow faster simulations and optimizations with varying values of geometrical parameters and subsequently can be implemented in high-level circuit and system design optimizations.
This thesis proposes a novel technique to develop combined neural network and pole-residue-based transfer function models for parametric modeling of EM behaviors of microwave components. In this technique, neural networks are trained to learn the relationships between pole/residues of the transfer functions and geometrical parameters. After the proposed modeling process, the trained model can be used to provide accurate and fast predictions of the EM behavior of microwave components with geometrical parameters as variables.
An advanced pole-residue tracking technique is proposed to exploit sensitivity information to solve the challenges of pole-residue tracking especially when the amount of training data are reduced and/or the geometrical step sizes between the data samples are enlarged. The proposed technique takes advantages of sensitivity information to increase of the orders of the transfer functions and ultimately form transfer functions of a constant order over the entire region of geometrical parameters. The proposed technique addresses the challenges of pole-residue tracking when training data are limited.
As a further advancement, we introduce EM sensitivity analysis into the pole-residue-based neuro-transfer function modeling technique. The purpose is to increase the model accuracy by utilizing EM sensitivity information and to speedup the model development process by reducing the number of training data required for developing the model. By exploiting the sensitivity information, the proposed technique can further speed up the model development process over the existing pole-residue parametric modeling method without using sensitivity analysis. The proposed parametric modeling techniques in this thesis are demonstrated by several microwave examples.