Learning-Based Rich Scattering Channel Estimation for U6G FDD XL-MIMO Systems

The extremely large-scale multiple-input multiple-output (XL-MIMO) under the upper 6 GHz (U6G) frequency band shows great potential for 6G wireless communications. However, deploying an extra-large array inevitably induces near-field effects, and the larger wavelength than millimeter waves makes buildings and vehicles effective scatterers in dense urban scenarios—resulting in complex multi-cluster, rich-scattering channels. These conditions pose challenges for downlink channel estimation in frequency division duplex (FDD) mode, as near-field research under U6G remains limited. In this letter, we target such rich-scattering environments by modeling a realistic multi-cluster multi-path channel and proposing an efficient deep learning–based estimator that exploits the spatial reciprocity between uplink and downlink. The uplink channel is visualized as a Cartesian-domain image, and an object detection network (CenterNet) locates the user and scatterers in a single inference step. To further reduce downlink pilots, a beamforming strategy is introduced for gain estimation. Numerical results demonstrate that the proposed FDD downlink estimation scheme achieves high accuracy and low overhead, providing a practical solution for near-field XL-MIMO under rich-scattering U6G conditions.