Recurrent Neural Network Modeling of Radio Frequency Amplifiers for System-Level Simulation and Design

We report a recurrent neural network (RNN) architecture with embedded static features for the prediction of the response of a power amplifier (PA) in the time domain. We show prediction of PA behavior across several signal amplitudes and bandwidths, with an average normalized mean squared error (NMSE) of −32.7 dB for transmission and −31.9 dB for reflection. In this way, we provide software-defined digital twin (DT) versions of hardware radio frequency (RF) amplifiers and further show the use of multiple DTs connected via a novel nested directed graph framework to form a digitally cascaded RF amplifier chain with applications to system level modeling. Current performance for a cascaded pre-amplifier and amplifier setup yields an average NMSE of −23.0 dB.