Enhancing Digital Predistortion Performance Under Load Mismatch Using a VSWR Generative Neural Network Simulator

This work introduces a VSWR generative neural network (GNN) simulator designed to simulate VSWR conditions as experienced by a power amplifier. The VSWR GNN simulator is trained using 20 recordings of the power amplifier (PA) and simulates the distortion caused by a given load using the 50-Ohm recorded signal. The generated data is then used to train a neural network-based digital predistortion (DPD) model, resulting in a DPD model that efficiently linearizes the PA across different loads. The performance of the proposed model is compared to a model trained solely on 50-Ohm data, using a 20 MHz 802.11ac WiFi signal on a CMOS PA, with an average power of 8 dBm and at a center frequency of 1 GHz. The proposed model achievs an average EVM improvement of 2 dB over the 50-Ohm model. This enhancement is accomplished through data synthesis during the training phase, without increasing model complexity.