Neural network electrothermal modeling approach for microwave active devices

Abstract

This article presents an artificial neural network (ANN) approaches for small‐ and large‐signal modeling of active devices. The small‐signal characteristics were modeled by S‐parameters based feedforward NN models. The models have been implemented to simulate the bias, frequency and temperature dependence of measured S‐parameters. Feedback NN based large‐signal model was developed and implemented to simulate the drain current and its inherent thermal effect due to self‐heating and ambient temperature. Both small‐ and large‐signal models have been validated by measurements for 100‐μm and 1‐mm GaN high electron mobility transistors and very good agreement was obtained.

Small signal behavioral modeling technique of GaN high electron mobility transistor using artificial neural network: An accurate, fast, and reliable approach

“…To draw a comparison with the proposed algorithm, in terms of accuracy and convergence rate, two well‐known algorithms known as steepest conjugate gradient (SCG) and Bayesian regularization (BR) is also used for training the small signal behavioral model for GaN HEMT based on cascaded architecture with the same set of network parameters as shown in Tables and , respectively.…”

“…An increase in the published literature is itself an evidence of the significance of machine learning techniques in device modeling . artificial neural network (ANN), one of the learning techniques, has been emerged as a powerful tool for device behavioral characterization and modeling and accommodates all the features of the machine learning . The advantage of ANN is that it can model highly nonlinear complex relations without even requiring explicit mathematical representations.…”

“…Owing to higher stability and fast convergence, the proposed GaN HEMT modeling technique, both the architectures are trained using Levenberg‐Marquardt (LM) algorithm. The conventional neural network‐based model mainly utilizes MLP architecture and model the device in form of magnitude and phase (in degrees) or using real and imaginary parts of S‐parameters . Real and imaginary parts of S‐parameter‐based model ensure the smoothness of the data and thereby increase the efficiency of the model.…”

“…[30][31][32][33] artificial neural network (ANN), one of the learning techniques, has been emerged as a powerful tool for device behavioral characterization and modeling and accommodates all the features of the machine learning. [34][35][36][37][38][39][40][41][42][43][44] The advantage of ANN is that it can model highly nonlinear complex relations without even requiring explicit mathematical representations. ANN is generally based on the concept of empirical risk minimization; therefore, local minima can also appear instead of global minima in some cases.…”

“…The conventional neural network-based model mainly utilizes MLP architecture and model the device in form of magnitude and phase (in degrees) or using real and imaginary parts of S-parameters. 35,[39][40][41][42] Real and imaginary parts of S-parameter-based model ensure the smoothness of the data and thereby increase the efficiency of the model. Unlike the conventional model, the present approach uses the concept of normalization for both magnitude and phase of two port S-parameters at preprocessing stage which seems to be a vital step obtaining better generalization capability and subsequently expediting convergence of the model.…”