Advanced Neural-Based Model Generation And Extrapolation Techniques For Microwave Applications

Neural-based modeling techniques have been recognized as important vehicles in the microwave computer-aided design(CAD)area in addressing the growing challenges of designing next generation microwave device,circuits and systems.Artificial neural network models can be trained to learn electromagnetic(EM)behavior of passive and active components/circuits.The trained neural network models can be used in high-level circuit and system design allowing faster simulation and optimization including EM and physics effects in components.The purpose of this thesis is to develop advanced neural-based model generation and extrapolation techniques for microwave applications.The proposed techniques take advantage of the high-efficiency of automated model generation algorithm,the cost-effective concept of knowledge-based neural network and the gener-alization capability of extrapolation techniques,to achieve reliable models for microwave applications.To further speed up neural modeling process,an automated knowledge-based neu-ral network model generation method using a new adaptive sampling technique for mi-crowave applications is firstly proposed.The proposed method integrates all the subtasks involved in knowledge-based neural modeling,thereby facilitating a more efficient and automated model development framework.The new adaptive sampling technique incor-porates interpolation techniques to determine the additional training samples required and their location in model input space.In this way,the proposed method can improve the ef-ficiency and reduce the expense of knowledge-based neural model development.For different modeling problems,the mapping structures in knowledge-based mod-els should be different.We propose a unified automated model structure adaptation al-gorithm for knowledge-based modeling using l1 optimization to automatically determine the type and topology of the mapping structure in a knowledge-based model.A new unified knowledge-based model structure to encompass various types of mappings is pro-posed.Using the distinctive property for feature selection of l1 optimization,the proposed method can automatically distinguish whether a mapping is needed or not and whether a mapping is linear or nonlinear.It is a more flexible and systematic technique and can further speed up the knowledge-based neural model development.As a further advancement,we propose an advanced multi-dimensional extrapolation technique for neural-based microwave modeling to make the model can be more reliably used outside the training range.Grid formulation in the extrapolation region is intro-duced and the proposed extrapolation is performed over these grids.We present multi-dimensional cubic polynomial extrapolation formulation and propose to use optimization to obtain extrapolated values at grid points.By using the proposed extrapolation method,neural models become more robust and reliable when they are used outside the training range.The validity of the proposed extrapolation method is demonstrated by both EM optimization example and nonlinear microwave simulation examples.