A unified neural network-based approach to nonlinear modeling and digital predistortion of RF power amplifier

Digital predistortion (DPD) has proven to be an efficient method of linearizing power amplifiers (PAs). In recent years, the use of neural networks (NNs) for DPD has gained momentum. This development can be attributed to the general attention NNs have gotten in recent years, but more importantly, to their ability to find efficient solutions to problems that either have no explicit solution, or the existing solution is sophisticated. Previously, DPDs were designed using one of the two main methods that originated from iterative control: direct learning architecture (DLA) and indirect learning architecture (ILA). Today, NN-based DPDs are using a new tool, but the same old architectures. In this article, we used the ability of an NN to break from the classic ways and come up with a new architecture, where a single NN models the entire system. As proof of concept, a 6-W single-transistor GaN PA, amplifying a 5G test signal, is linearized using the proposed method.