Artificial Neural Network Nonlinear Transistor Behavioral Models: Structure and Parameter Determination Process Based on the Cardiff Model

This article introduces a novel artificial neural network (ANN) structure determination process based on the Cardiff model (CM), to determine ANN-based transistor nonlinear behavioral models. By relating the CM formulation and coefficients to the Taylor series expansion of the ANN model, a novel approach for determining the required values of a fully connected cascaded (FCC) ANN structure has been formulated. The proposed method provides the chance to escape from the possible time-consuming ANN determination process. Experiments proved that the proposed ANN models using the determination method can provide accurate prediction for the behavior acquired from load-pull characterizations of a Wolfspeed 10-W packaged gallium nitride (GaN) high electron mobility transistor (HEMT) simulation at 3.5 GHz, and a dense load-pull measurement of WIN NP $12 4 \,\, \times 75~\mu $ m GaN HEMT at 20 GHz, with normalized mean square error (NMSE) levels lower than −40 dB.