Advanced Space Mapping And Artificial Neural Networkbased Techniques For Microwave Device Modeling

Microwave device modeling is an important topic in microwave area.This thesis focuses on developing advanced space mapping and artificial neural network(ANN)-based techniques for challenging microwave device modeling,addressing both microwave active and passive component modelings.Specifically,this thesis develops an advanced space mapping-based technique in combination with knowledge-based neural network(KBNN)for modeling gallium nitride(GaN)high-electron-mobility transistor(HEMT)devices.This thesis also develops advanced ANN-based transfer function mapping techniques for parametric modeling of microwave passive components.GaN HEMTs are important for next-generation wireless communication systems and microwave power device applications.However,GaN HEMT model development can be time-consuming as the devices exhibit strong trapping effects,which often require very sophisticated model.As a consequence,a fast modeling approach for GaN HEMT devices with trapping effects is of significance for high-reliability microwave circuit design.In the first part of this thesis,we propose a novel space mapping modeling technique for GaN HEMTs with trapping effects.The proposed space mapping technique develops separate mappings for different branches inside the existing device model,such that different behaviors(i.e.,trapping effects and frequency dispersion)in GaN HEMTs can be mapped separately.Through supervised learning methods,each mapping module is systematically developed to overcome the gap between each internal branch and each set of target data,accelerating the process of model development.The KBNN model is proposed for characterizing drain current under DC,and it is used as part of the trapping mapping development.The proposed space mapping technique allows fast and systematic model development for GaN HEMTs with trapping effects to achieve a GaN HEMT large-signal model with satisfactory accuracy.Once developed,the large-signal model can be used in high-performance microwave circuit design,covering DC,PIV,S-parameter,HB,and load-pull simulations.Neuro-transfer function(neuro-TF)approaches have become popular in parametric modeling for EM behavior of microwave components.Neuro-TF model combines neural networks and a transfer function,where transfer function is used as space mapping between the neural network and the EM responses.In the second part of this thesis,we propose a novel hybrid-based neuro-TF technique to address the discontinuity issue and the non-smoothness issue of the poles/residues when the geometrical variations of the EM structures become large.In the proposed technique,we systematically combine both pole-residue and rational formats of the transfer functions in the neuro-TF model.The proposed technique automatically identifies the poles/residues that are smooth-continuous and the poles/residues that have the discontinuity and non-smoothness issues.The proposed technique converts the poles/residues that have those issues into the coefficients of the rational-based transfer function and remains the smooth-continuous poles/residues in the pole-residue format of the transfer function.The proposed technique can obtain good model accuracy in challenging applications of large geometrical variations,addressing the discontinuity and non-smoothness issues.In the third part of this thesis,we propose a novel decomposition technique for the rational-based neuro-TF approach to address the high-sensitivity issue when the geometrical variations become large,where the sensitivity means the sensitivity of EM response with respect to the coefficients of the rational-based transfer function.We propose to decompose the original rational-based neuro-TF model with high-order transfer function into multiple sub-rational-based neuro-TF models with much lower-order transfer function.We then reformulate the overall model as the combination of the sub-neuro-TF models.The proposed technique automatically determines the number of sub-models and the transfer function order for each sub-models.The proposed technique decreases the sensitivities of the overall model response with respect to the coefficients of the transfer function in each sub-models,leading to robust training of the overall model.The proposed modeling technique can achieve good model accuracy in challenging applications of large geometrical variations,addressing the high-sensitivity issue.