Learning When to Learn: Distortion-Aware GNN Precoding for Multi-Carrier Large Antenna Systems with Adaptive Precoder Selection

In large-antenna systems, operating power amplifiers (PAs) close to saturation is key to achieving high energy efficiency. However, this operating regime inevitably introduces nonlinearities, whose resulting distortion can coherently combine at the user location and degrade system performance. Distortion-aware precoding techniques have been proposed to send this distortion in non-user directions. These distortion-aware precoders are typically designed under the assumption of a frequency flat channel. In contrast, we assume a multi-carrier, single-user downlink system that experiences frequency-selectivity. In multi-carrier systems, the non-linearity of the PA produces intermodulation products, creating inter-carrier interference so that the precoding problem becomes coupled across subcarriers. To tackle this, we propose a graph neural network (GNN)-based joint precoding method that optimizes precoding vectors across all subcarriers while accounting for distortion. Simulation results indicate that this joint GNN precoder achieves stable rates across a wide range of channel conditions, outperforming classical maximum ratio transmission and per-subcarrier GNN approaches, albeit at increased computational complexity. Additionally, we show that the need for such a method depends strongly on the channel conditions. As such, we propose a convolutional neural network-based classifier that selects the appropriate precoding method based on the current channel conditions, while balancing the achievable rate and complexity.