GRP-0275 Graph Attention Network based Downlink Channel Prediction using in Frequency Division Duplexed NextGen Networks

GRP-0275 Graph Attention Network based Downlink Channel Prediction using in Frequency Division Duplexed NextGen Networks

https://www.kennesaw.edu/ccse/events/computing-showcase/fa25-cday-program.php

In Frequency Division Duplex (FDD) 5G networks, downlink channel state information (CSI) must be estimated at the user equipment (UE) and fed back to the base station, a process that requires frequent CSI-RS transmission and uplink feedback, resulting in high overhead and energy consumption. This research proposes a novel base-station–centric framework that predicts the downlink channel matrix directly at the gNB, eliminating the need for continuous CSI-RS–based estimation at the UE. By leveraging uplink channel observations, geometric environment features, and learned mappings between uplink and downlink channel relationships, our model reconstructs the downlink MIMO channel with high fidelity. The system integrates ray-tracing-based dataset generated via NVIDIA Sionna, combined with graph attention networks and transformer encoders to capture spatial and temporal channel dependencies. We infer the downlink channel without per-slot CSI-RS transmission based in environment geometry and uplink transmission signals, significantly reducing signaling overhead while maintaining beamforming performance. This work enables a shift toward AI-assisted FDD systems where proactive channel prediction replaces periodic downlink probing, contributing to greener and more efficient 5G/6G networks.