Compact Modelling Of RF And Microwave High Power Gallium Nitride Transistor

Gallium nitride(GaN)is the third-generation semiconductor material with the property of wide band gap,high heat capacity,good heat conductivity,and high electron mobility.The ultimate goal of advanced semiconductor process is the chip design,and compact models are the bridge linking the semiconductor process and chip design.This thesis focuses on the compact modeling of GaN HEMT,and the original contributions are summarized as below:1.Optimization based small signal equivalent circuit parameter extraction method of GaN HEMT:We proposed a novel intrinsic and extrinsic Y-parameter loss function,which is robust to measurement uncertainty and noise.The Y-parameter loss function is carefully designed based on the ? structure of the equivalent circuit,thus the error calculation is much more efficient compared with that of the conventional S-parameter loss function.We choose artificial bee colony as the optimizer with the parameter boundaries provided by the conventional open-short or cold-FET method,so that the final solution can achieve excellent accuracy without losing physical meaning.This algorithm is highly adaptive to semiconductor processes,and can always achieves good results when other extraction methods fail,thus could be an important complement to the existing extraction methods.2.Dimension reduction method for the dispersion modeling of semiconductor devices:This method integrates empirical functions with multi-variable Taylor series.The dimensions with strong nonlinearity are represented by conventional empirical functions,while the dimensions with weak nonlinearity are expanded by Taylor Series.For example,the dependence of terminal voltages of the channel current of GaN HEMT are described by empirical functions,while the dispersive dimensions of traps and thermal are represented by Taylor expansion.The final model is completely an analytical function,thus the simulation convergence and model extrapolation ability are guaranted.This thesis provides the concrete theory with detailed model extraction procedure,which can achieve the automation of model building.This dimension reduction method is purely mathematical which is applicable to the compact modeling of various semiconductor devices.To verify its generality,an extra example of the thermal modeling of a LDMOS is given.This method is proposed to incorporate in the commercial device modeling software IC-CAP by KeySight Technology,which shows good academic and industrial value.3.Feedforward artificial neural network based electrothermal model for GaN HEMTs with dynamic trapping effects consideration:This large signal model is mainly composed of nonlinear current and charge sources,which incorporates dynamic trapping and thermal effects.We have done extensive investigation and analysis on the lag effects due to the traps,and the trapping states are successfully identified from the pulsed I-V measurements.These dispersive states are essential to the training of the artificial neural network based current source.Meanwhile we proposed three sub-circuits to mimic the dynamic process of electron trapping,emission and heat propagation.We also proposed an integral neural network structure for the charge source construction.The charge sources with thermal dependence can be trained from the nonlinear capacitances.Compared with the conventional nonliear capacitance model model,charge models significantly improves the simulation convergence and efficiency,and accelerates simulation speed of large-scale microwave circuits.Finally a GaN HEMT provided by Dynax was employed to validate the proposed modeling methd,and good results had been achieved in terms of I-Vs with various trapping and thermal states,multi-bias S-paramters,and large signal characteristics.4.Dimension-Reduced artificial neural network for the compact modeling of semiconductor devices:We proposed a dimension-reduced artificial neural network(DRANK)architecture,which integrates the Taylor series with fully connected neural network.For a multi-dimensional input and output-mapping problem,the dimensions with strong nonlinearity(e.g.terminal voltages)are described by fully connected artificial neural network,while the rest weakly nonlinear dimensions(e.g.traps,temperature)are expanded by Taylor series.This hybrid architecture could significantly reduce the necessary training data compared with the conventional fully connected neural network,for instance,if we only have two data points for a specific dimension,we would definitely expect an over-fitted neural network,however,linear regression(first order Taylor series)can result to a robust model.The DRANN utilizes the combinations of several lower dimensional neural network to approximate a higher dimensional neural network.We have employed a GaN HEMT from Dynax to do the verification,and the experiment shows that our method can accurately capture the traps and thermal effects while overfitting problem is avoided.To demonstrate the generality of DRANN,we also provide a LDMOS thermal modeling example.The DRANN architecture can be applied to the compact modeling of various semiconductor devices or system level modeling.