Modeling HBT Characteristics Based On Improved Wiener-type Neural Network

The research of device modeling methods has become one of the important hot spots in the field of microwave design.Heterojunction bipolar transistor(HBT)has a series of advantages such as short crossing time,high cutoff frequency,and large current gain.It has been widely used in wireless communication,satellite communication,and Internet of Things.One of the most important issues is the accurate modeling of HBT.Artificial neural networks have been widely used in device modeling because they are not only efficient but also guarantee the accuracy of the model compared with traditional modeling methods.Therefore,the neural network approach is introduced for the characterization of HBTs.In this thesis,a novel Wiener-type dynamic neural network modeling method applicable to the modeling of HBT is proposed and it is used to research the characteristics of HBT.The method first considers the transmission characteristics of HBT,and determines the structure of a modified Wiener-type dynamic neural network model applicable to HBT by vector fitting of the H-parameter.Secondly,in order to train this improved structure,we also propose an improved training algorithm and give the detailed computational procedure and flowchart of this improved training algorithm.Among them,we consider the transmission characteristics as well as the amplification characteristics of the HBT,use the H-parameter instead of the Yparameter for small-signal analysis and derive the DC and small-signal calculation formulas for HBT calculations.Finally,the proposed novel Wiener-type dynamic neural network method is used in the research of DC characteristics and small-signal characteristics of HBT,which is verified by two arithmetic examples.In the case simulation,we have built DC and small signal models by training GaAs HBT and InGaP/GaAs HBT,and both cases can accurately characterize the DC and small signal characteristics of the devices.The training error and testing error of the trained model can be controlled within 2%.In this thesis,the Wiener-type dynamic neural network algorithm and calculation formula are improved to making it suitable for the research of the characteristics of HBT.The proposed novel Wiener-type dynamic neural network does not require detailed parameters and internal information of HBT devices to model HBTs with different substrates.The simulation results demonstrate the feasibility and accuracy of the proposed improved Wiener-type dynamic neural network modeling method,which can also be used in computer-aided design and other fields,and provides a certain theoretical basis and reference for the modeling research of HBT.