Large-Signal Modeling of GaN HEMTs Using Hybrid GA-ANN, PSO-SVR, and GPR-Based Approaches

Abstract

This paper presents an extensive study and demonstration of efficient electrothermal largesignal GaN HEMT modeling approaches based on combined techniques of Genetic Algorithm (GA) with Artificial Neural Networks (ANN), and Particle Swarm optimization (PSO) with Support Vector Regression (SVR). Another promising Gaussian Process Regression (GPR) based large-signal modeling approach is also explored and presented. The GA-ANN addresses the typical problem of local minima associated with the backpropagation (BP) based ANN. The GA successfully aids in the determination of optimal initial values for BP-ANN and enables it to find a unique optimal solution after subsequent of iterations with higher rate of convergence. This is also achieved using PSO-SVR with lower optimization variables. The developed modeling techniques are demonstrated and used to simulate the gate and drain currents of a 2-mm GaN device. All the models are relatively simple, practical, and easy to implement. The gate and drain currents models are embedded in an equivalent large-signal circuit's model and built in Advanced Design System (ADS) software. The implemented model is validated by large-signal measurements and very good fitting results have been obtained. The model also showed an accurate simulation for a nonlinear power amplifier with very good computational speed and convergence.

Multilayer perceptron–random forest based hybrid machine learning–neural network model for GaN high electron mobility transistor's parameter estimations

“…It is to be noted that training time cannot be used as a standard metric for comparing the different models reported in the literature. For a given model, the training time depends on the dataset, which is not standard across the different papers used for comparison, 11,62,64 which have considered a different set of the number of inputs, and the test cases. For example, in case of Sample A, a total of 237 600 data points are considered for power metrics estimation, for which ∼10 s is recorded for model training.…”

“…Owing to several advantages like fast learning and generalizing capability, strong robustness to the noise in training data (using the attribute‐value pair schema), and ability to sustain long training in case of heavy data, large nodes, and complex functions; ANNs are growing choices for problems based on transistor modeling. Table 1 compiles an exhaustive list of works discussing the application of different NN based techniques for device modeling 51–71 . The direct feature engineering in ML and NN algorithms enables straightforward interpretation and adaptability of such techniques to varying device parameters, which gives an edge over other methods and is the focus of this work.…”

Role of Machine Learning in Rapid Modeling of RF Devices: VO2 RF Switch Modeling as a Case Study

“…The neurons in both the hidden layers varied from 15 to 25 with step size of 1 whereas the activation function at all three layers can be chosen from the list stated in Table 1. The basis of training the neural network model involves computation of weighting matrix and bias, and optimal selection of number of neurons and associated activation function at each layer so that the mean square error (MSE) function, in (10) , is minimized. Here, N is sample size, y meas represents the measured value and y pred is the predicted value.…”

Equivalent‐circuit extraction for gallium nitride electron devices: Direct versus optimization‐empowered approaches

“…Over the last half century or so, many efforts have been devoted to the equivalent‐circuit extraction for modeling the scattering S‐parameter measurements of active electron devices 1–21 . Although the frequency‐dependent behavior of the scattering parameters can be easily and quickly reproduced by using alternative representations (e.g., artificial neural networks 22–26 ), the calculation of a small‐signal equivalent‐circuit model is essential as this circuit can be utilized as a foundation for developing large‐signal 27–32 and noise 33–38 models. Typically for the field‐effect transistors (FETs), the challenging task of extracting the small‐signal equivalent‐circuit model is accomplished by using direct extraction techniques based on the well‐known “cold” approach 4,5,10,12,13 .…”