Neural-Based Large-Signal Device Models Learning First-Order Derivative Parameters for Intermodulation Distortion Prediction

A detailed procedure to learn a nonlinear model together with its first-order derivative data is presented. Two correlated multilayer perceptron (MLP) neural networks providing the model and its first-order derivatives, respectively, are trained simultaneously. Applying this method to FET devices leads to nonlinear models for current and charge fitting derivative parameters. The training data is the bias-dependent equivalent circuit parameters extracted from S-parameter measurements. The resulting models are suitable for both small-signal and large-signal analyses, in particular for intermodulation distortion prediction. Examples for power amplifier simulations of power transfer, efficiency and intermodulation distortion performances are presented.