Comprehensive Investigation and Comparative Analysis of Machine Learning-Based Small-Signal Modelling Techniques for GaN HEMTs

Abstract

A number of machine learning (ML) algorithm based small signal modeling of Gallium Nitride (GaN) High Electron Mobility Transistors (HEMTs) have been reported in literature. However, these techniques rarely provide any inkling about their suitability in modeling GaN HEMTs under varied operating conditions. In this context, this paper thoroughly investigates various ML based techniques and identifies their suitability for specific application scenarios. At first, an array of commonly employed modeling techniques based around Artificial Neural Network, RANdom SAmple Consensus, Support Vector Regression, Gaussian Process Regression, Decision Tree, and Genetic algorithm assisted Artificial Neural Network are used for development of modeling framework to exploit the bias, frequency and geometry dependence on S-parameter based outputs. Subsequently, the ensemble techniques namely Bootstrap aggregating, Random Forests, Extremely Randomized Trees, AdaBoost, Gradient Tree Boosting, Histogram-based Gradient Boosting, and Extreme Gradient Boosting are also explored to understand the capability of these algorithms in the development of GaN HEMT small signal models. Thereafter, an exhaustive analysis of bias and variance is carried out to figure out the most appropriate algorithms for specific applications. The discrepancies during model development are removed by tuning the hyperparameters of the respective models using Random search optimization with 5-fold cross validation technique. Post tuning, the models are evaluated in terms of generalization capability, Advanced Design System compatibility, computational efficiency, training and simulation time, models' capacity and parameters'tuning time.

Accurate, Efficient and Reliable Small-Signal Modeling Approaches for GaN HEMTs

“…The second phase, which enables the POA to have better exploitation ability is mathematically represented as (19), where x P2 i,j , R, R.(1-t MNOI )-mainly responsible for exploitation, and t denote updated position of ith pelican in the jth dimension, constant number (R = 2), nearby area around x i,j and current iteration, respectively. POA updates the positions of each member according to stages 1 and 2 and repeats this process until the termination condition is met.…”

“…Step 4: to get to the next iteration, as explained earlier, positions of pelicans are updated using two-stage formationpelicans' movement towards prey and winging behavior according to ( 18)- (19).…”

Particle swarm optimization‐XGBoost‐based modeling of radio‐frequency power amplifier under different temperatures

“…Computer‐aided design (CAD) has been widely used as it can complete the modeling of RF PA with lower cost and difficulty 6 . Artificial neural network (ANN) is commonly used in modeling due to its excellent ability to express nonlinear functions 7 .…”

PSO‐GA‐SVR model for S‐parameters of radio‐frequency power amplifier under different temperatures

“…Currently, radio‐frequency (RF) power amplifier (PA) has been widely used in wireless communication and radar 1,2 . Designing a high‐performance RF PA has become a widespread demand, 3 which increases the demand for modeling RF PA 4 . Computer‐aided Design (CAD) has been closely combined with the modeling of RF PA to quickly and efficiently assist high‐precision modeling of RF PA, thus assisting circuit design 5,6 .…”