Deep Neural Network-Based Digital Predistorter for Doherty Power Amplifiers

Abstract

In this letter, measured adjacent channel leakage ratio (ACLR) results using a GaN Doherty power amplifier will show that for less than 2000 coefficients, sigmoid activated deep neural network (DNN)-based digital predistorter (DPD) outperforms rectified linear unit (ReLU) activation by up to 2 dB even when the number of layers of the network is increased. When the number of coefficients exceeds 2000 ReLU outperforms sigmoid activation with an improvement of up to 3–4 dB in ACLR suppression. Furthermore, to achieve an ACLR level of −54 dBc or better, the number of coefficients required to implement the DNN-DPD can be reduced by a factor of 150 when using ReLU rather than sigmoid activation.