Accurate and Efficient Modeling of FET Cold Noise Sources Using ANNs

Abstract

Accurate and efficient models of the available output noise temperature of a field-effect transistor (FET) cold noise source were developed using artificial neural networks (ANNs). Radial basis function and backpropagation (BP) architectures were used for this application, and the models were obtained from measured data of a gallium arsenide (GaAs) metal-semiconductor FET (MESFET) with a gate width and a 0.25-mum gate length. Two-input one-output ANNs were developed for incident noise temperature versus load reflection coefficient, as well as incident noise temperature versus load impedance for different percentages of training data. The most accurate model was a BP network that used 75% of the measured data and the Levenberg-Marquardt algorithm for training and was for incident noise temperature versus load impedance with average relative error (ARE) and maximum local relative error (MLRE) values of 0.1439% and 1.1544%, respectively. The most efficient model was a BP network that used only 42% of the measured data and the resilient backpropagation algorithm for training and was for incident noise temperature versus load impedance with an ARE of 0.1940% and an MLRE of less than 5%. To further validate the modeling process, a set of measured data from a GaAs MESFET with a gate width and a 0.25- gate length was used with the architecture of the previous models to develop new models. All results demonstrated that ANNs can accurately and efficiently predict incident noise temperature of cold noise sources.