DEVELOPMENT AND OPTIMIZATION OF ML BASED COMPREHENSIVE MODELLING FRAMEWORK FOR GAN HEMTS

Abstract

Radio Frequency (RF) Power Amplifier (PA) is one of the most pivotal constituents of any wireless transceivers. However, continual advancements and ever-increasing complexity in the wireless communication technologies demand frequent innovations in the design of RFPAs. The quality of the designed RFPAs are generally evaluated based around two basic figures of merits namely efficiency and linearity. Thus, the RFPAs should provide maximum power and efficiency while maintaining highly linear operation. In literature, two primary PA design mechanisms, namely measurement- and modeling-based techniques have been extensively utilized. Each class of technique has pronounced merits, limitations and applications. However, owing to the seamless integration ability of the modeling-based techniques with Computer-Aided Design (CAD) tools, they are increasingly becoming more popular. The design and innovation in RFPAs are excessively contingent on the measurement facilities and the Large Signal Models (LSMs) of transistor devices. At present, Gallium Nitride (GaN) High Electron Mobility Transistor (HEMT) technology is regarded as an optimal microwave transistor technology for the design of RFPAs in advanced RF/microwave and high power switching applications. This is due to their attributes namely high energy bandgap, high saturation velocity, high electron mobility, exceptional thermal behavior and high breakdown field. Furthermore, GaN HEMTs manifest high power density, thus a smaller size device can be used to sustain a high power demand. It also implies reduced lower capacitances and lower combining losses in the design of RFPAs and Low-Noise Amplifiers (LNAs). At this point, it is essential to mention that, in general, the available LSMs of GaN HEMTs are very specific and therefore not readily useful for broad range of PA designs. Therefore, there is a pressing requirement to develop accurate, reliable, efficient and robust LSMs of GaN HEMTs which can be readily incorporated in CAD tools. Nevertheless, Small-Signal Model (SSM) development is the first step in pursuit of developing accurate and efficient LSMs. But, both SSMs and LSMs of GaN HEMTs are essential for the design of accurate, efficient and reliable GaN HEMT based RFPAs. Apparently, various modeling schemes have been exploited to develop SSMs and LSMs for GaN HEMTs, however, usually, they are classified into three main groups, which are physics-based, Equivalent Circuit (EC) and Behavioral Modeling (BM) frameworks. This thesis is originated in response to the scientific and technical challenges in EC and BM frameworks for GaN HEMTs at high frequency applications. Among these challenges, the major focuses are on the development of SSMs for GaN HEMTs, which are simple, accurate, computational and time efficient, reliable, scalable, and CAD adaptable. Furthermore, special attention is given to develop SSMs, which manifest strong interpolation and extrapolation abilities. The developed SSMs are then utilized to realize the eventual LSMs for GaN HEMTs. In order to develop SSMs and LSMs for GaN HEMTs, which possess the above-mentioned characteristics, in this thesis, Machine Learning (ML) based approaches have been explored and utilized because of their superior learning, prediction, and extrapolation abilities. However, it is pertinent to state that the ML based modelling of GaN HEMTs is still in its early exploration phase, and various issues related to this type of modelling are unexplored and not thoroughly discussed in literature. It is therefore, in this thesis, an extensive appraisal and analysis of ML and optimization based small-signal and large-signal modelling for GaN HEMTs have been presented. In the first part of this thesis, a detailed comparative analysis of EC based accurate, robust and efficient SSM parameter extraction methodologies for GaN-on-Diamond HEMTs has been demonstrated. For this, initially, a Scanning- Based Systematic (SBS) model parameter extraction approach is developed and applied on GaN-on-Diamond HEMTs. Thereafter, marine predators algorithm, pelican optimization algorithm and tunicate swarm algorithm, the recently developed Optimization Algorithms (OAs), based hybrid extraction methodologies have been developed and applied on the same GaN HEMTs. Finally, a detailed comparison of OAs and SBS modelling schemes by using SBS extraction approach as a benchmark in terms of reliability, accuracy, convergence behavior, complexity, execution time, and scalability is provided and thoroughly discussed. Accurate, efficient and CAD compatible small-signal behavioral models for GaN HEMTs using Artificial Neural Network (ANN), Support Vector Regression (SVR) and Gaussian Process Regression (GPR) based ML techniques have been developed, validated and discussed in the subsequent part of this thesis. These ML based approaches have been applied on many GaN HEMTs devices grown on Silicon (Si), Silicon Carbide (SiC) and Diamond substrates. Furthermore, a meticulous evaluation of ANN algorithms implemented in MATLAB, Python (using Keras, PyTorch and Scikit-learn) and R (using H2O) for small-signal behavioral modelling of GaN HEMTs has been presented. To establish the appropriateness of software environments in distinct application settings, the developed models are examined on a range of metrics namely behavior on the unseen data, training and prediction speed and ADS adaptability, and software environments are surveyed for support and documentation, user-friendly interface, simplicity in the model development procedure, open-access and cost. Optimization of the hyperparameters of ML algorithms is vital to realize the best possible models. In this context, hybrid optimized ML algorithms namely Genetic Algorithm (GA) and Particle Swarm Optimization (PSO) assisted ANN, PSO and RSO assisted SVR, RSO assisted GPR, and RSO assisted various tree-based models are explored and developed. Finally, the developed models are evaluated on many regression tests to identify the most fitting ML algorithms for particular applications. Finally, the last part of the thesis presents ML based CAD adaptable advanced models and applications. Initially, GPR, GA-ANN and RSO-Decision trees based SSMs for GaN HEMTs are developed. Then, the integration of these developed models with ADS are presented by inserting the developed ML based models into a design of class-F PA. Subsequently, to examine the accuracy of the models, stability and gain tests of the GaN HEMT based amplifier configuration are performed. Thereafter, using the developed SSMs, a joint EC-behavioral LSM for a GaN HEMT is developed and presented. The intrinsic drain and gate currents are modelled using GA-ANN, PSOSVR and GPR based approaches. These current modelling approaches are compared in terms of simplicity in the model development stage, computational efficiency, accuracy and required time to simulate the currents. At last, LSM validation and realization using GA-ANN based approach are demonstrated on a design of an inverse class-F PA.