High-frequency Noise Model Of SiGe HBT Based On Artificial Neural Network

Heterojunction bipolar transistors(HBTs)have become the first choice of modern and next generation wireless communication systems in millimeter wave band because of their excellent high frequency performance and their compatibility with silicon-based CMOS technology.The key to realize the design and optimization of high frequency circuit based on HBT is to use effective and accurate HBT model to simulate various performances of analog circuit.At present,the accuracy of small signal equivalent circuit model of active devices,which is widely used in microwave and millimeter wave circuit simulation design software,it is limited by the test frequency range of modeling measurement data and the bias conditions of the tested devices.The simulation capability of frequency interpolation and bias point epitaxy of HBT high frequency small signal equivalent circuit model is insufficient.With the increase of the working frequency,the complexity of the small signal equivalent circuit model of HBT is also increased,which increases the difficulty of accurate extraction of the parameters of the equivalent circuit model and the development cost and the development cycle of microwave and millimeter wave HBT circuit is also increased.In this paper,the effectiveness,accuracy and simplicity of Artificial neural network(ANN)in the modeling of active devices are utilized,and the ANN modeling of SiGe HBT high-frequency noise in millimeter wave frequency band is studied in the following three aspects.Firstly,it is considered that the core of ANN modeling of active devices is to train the neuron parameters of ANN model under a certain type or structure,so that it has the ability to characterize the nonlinear relationship between the working conditions and performance parameters of devices.In view of the strong generalization ability of ANN modeling technology,different active devices may be suitable for similar ANN type or topology mechanism and the same optimization training algorithm,so the key of ANN modeling of different active devices is to select accurate and reliable training data and test samples.Therefore,this paper studies the direct extraction method of the parameters of SiGe HBT high frequency small signal equivalent circuit model.Among them,on the basis of theequivalent circuit fully characterizing the non quasi-static effect of HBT,through the mathematical analysis of the Taylor expansion of the equivalent circuit admittance(Y)and impedance parameter(Z)models,the accurate extraction of the equivalent circuit parameters is realized,which lays the foundation for training the scattering(S)parameter ANN model established in this paper to provide accurate broadband S parameter simulation data.At the same time,in view of the accuracy of SiGe HBT high-frequency shot noise model of Transport and SPICE model is not enough,this paper presents an effective shot noise semi-empirical model,which can cover all bias conditions and can be used to verify the results of this paper built the four noise parameters of the ANN model based on knowledge representation device noise the precision of offset features provides a test sample.Secondly,by referring to the experience of ANN modeling of RF and microwave devices and using the generalization ability of ANN modeling,this paper selects the neural network structure of multi-layer perceptron and uses the genetic algorithm(GA)+Levenberg Marquardt back propagation(L-M BP)algorithm as the optimization training algorithm to establish the S-parameter ANN model of SiGe HBT.Among them,by introducing GA algorithm with global optimization ability,the problem that the training convergence speed of L-M algorithm of BP neural network is slow and easy to be limited to local extremum is solved.It is proved that the S-parameter ANN model has good interpolation ability in a wide frequency band and bias range.Therefore,the S-parameter ANN model can easily obtain the S-parameter data at any bias point and frequency range,and avoid the complicated S-parameter measurement process and the tedious equivalent circuit parameter extraction process.Using the high precision simulation data generated by the S-parameter ANN model as a priori knowledge,it provides a guarantee for improving the accuracy of the four noise parameter ANN modeling.Finally,this paper established a four noise parameter ANN model based on knowledge.The model type is multi-layer perceptron,and the training algorithm is GA + L-M BP.The purpose of building the model is to quickly and continuously characterize the bias and frequency dependence of SiGe HBT high-frequency noise in a wide frequency band and bias range,so as to avoid the huge demand for complex four noise parameter measurement in theanalysis and modeling of HBT high-frequency noise,and easy to obtain the shot noise data at any bias point and frequency.In order to verify the prediction accuracy of the knowledge-based ANN model with four noise parameters,this paper also presents an algorithm to extract the power spectral density of the noise current source based on the HBT two port equivalent noise model with four noise parameters.The experimental results show that the simulation results of the ANN model based on the four noise parameters and the semi empirical model based on the measurement have good consistency,which verifies the validity and accuracy of the model.In addition,compared with the four noise parameter ANN model without knowledge,the four noise parameter ANN model based on knowledge has obvious advantages in modeling efficiency and modeling accuracy.