A highly efficient approach to the development and implementation of a charge-controlled large-signal knowledge-based neural network model for HEMT devices.

摘要

Large-signal equivalent models are difficult to develop and implement into existing CAD tools presenting a significant challenge to their widespread use in engineering modeling systems. However, the creation of large-signal models is beneficial in helping designers accurately model nonlinear behavior in both passive and active components. Presented in this dissertation is a highly efficient approach to the development and implementation of a large-signal knowledge-based neural network model for a HEMT device. The development of unique expressions capable of modeling the various nonlinear relationships in a transistor, the novel method by which charge is calculated and the development of an innovative engine used to integrate the model into existing CAD tools is what makes this overall approach highly efficient and extremely accurate. The advantage of this approach is demonstrated in this work by comparison of measured and simulated DC I-V curves over a wide dynamic range, derivatives of DC I-V curves up to the 4th order, scattering parameter measurements at different bias points and power-out vs. power-in ratios along with power added efficiency of a 1500 μm HEMT device.; This work discusses the development of a backpropagation and a knowledge-based neural network engine capable of automating the construction, configuration, and training of neural network models. Through the use of the engine, a parameter file is generated from a pre-trained neural network that enables neural network models to reproduce specific functionality. These models replace selected components with neural network structures that can be dynamically configured and are implemented in Agilent Technologies' Advanced Design System (ADS) circuit simulator. These types of models significantly reduce optimization time in standard optimization and statistical design approaches that require repeated circuit simulations. In addition, they provide more accurate solutions than polynomial models and their size is more manageable when increasing the dimension of the model. Due to the non-specific nature of these models, they can be readily applied to any design. Designers can easily integrate these types of models into their designs and take advantage of the benefits neural networks offer.