Knowledge-based neural networks for microwave modeling and design

Neural networks have recently been introduced to the microwave area as a fast and flexible vehicle for microwave modeling, simulation and optimization. This thesis represents a new research direction in this area, i.e., incorporating microwave knowledge into neural network models. Two approaches utilizing functional and structural microwave knowledge are proposed in the thesis. They are used for efficient modeling of active and passive microwave components and for the development of libraries of microwave neural models.;In the first approach, a novel neural network structure, namely functional knowledge based neural network (KBNN), is proposed where microwave empirical or semi-analytical information is incorporated into the internal structure of neural networks. The microwave knowledge complements the capability of learning and generalization of neural networks by providing additional information which may not be adequately represented in a limited set of training data. Such knowledge becomes even more valuable when the neural model is used to extrapolate beyond the training-data region. A new training scheme employing the gradient-based l2 optimization technique is developed to train the KBNN model. The proposed technique can be used to model passive and active microwave components with improved accuracy, reduced cost of model development and less need of training data over conventional neural models for microwave design.;The second approach addresses the development of libraries of neural models for passive and active components, a task with a potential significance for many microwave simulators. Developing libraries of neural models is very costly due to massive data generation and repeated neural network training. A new hierarchical neural network approach is proposed, allowing both microwave functional knowledge and library inherent structural knowledge to be incorporated into neural models. The library models are developed through a set of base neural models, which capture the basic characteristics common to the entire library, and high-level neural modules which map the information from base models to the library model outputs. The proposed method substantially reduces the cost of library development through reduced need for data. collection and shortened time of training. The technique is demonstrated through transmission line and FET library examples.