A neural network characterization of a HEMT

Abstract

We report a new approach to describe the bias-dependent behavior of a HEMT by using a neural network, whose inputs are gate-to-source (Vgs) and gate-to-drain bias voltages (Vds). Using a conventional small-signal equivalent circuit, we characterized the HEMT's S-parameters measured at various bias settings, and obtained the bias-dependent values of the equivalent circuit elements. Through experiments, we found that a 5-layered neural network (composed of 28 neurons) is adequate to represent 7 bias-dependent intrinsic elements simultaneously. A "well-trained" neural network shows excellent accuracy.