AI-Based Beam Management for FR3 FDD MIMO via Online Channel Synthesis

H.-J. Moon, J. Park, C.-B. Chae, R. W. Heath Jr.
Published in IEEE Journal on Selected Areas in Communications, Jan. 2026

Abstract
Efficient channel estimation and beamforming in the upper mid-band spectrum pose a fundamental challenge for 6G base stations (BSs). This requires practical beam management techniques that account for the propagation characteristics of this frequency range and the increasing number of BS antennas. To address the channel estimation bottleneck in Frequency Division Duplex (FDD) systems, this paper proposes an integrated framework combining channel synthesis and AI-based beam management. We first introduce partial statistical reciprocity (PSR) by extending the concept of partial reciprocity into the statistical domain, enabling a novel model-based online channel synthesis method. The proposed approach uses uplink channel parameters collected during BS operation to generate synthetic downlink channels that reflect the spatial distribution of active user equipment (UE), enabling the creation of high-quality, site-specific datasets for AI training. Furthermore, we propose a joint optimization framework for scalable codebook design and downlink channel estimation. A U-Net Transformer model is trained to estimate full beam-domain channel state information (CSI) using a single snapshot of uplink CSI and feedback from codebook-based beam sweeping. The model iteratively updates the codebook via backpropagation to align with current UE distributions and channel conditions, while channel estimation is performed through forward inference. The proposed AI-based framework achieves accurate channel reconstruction with minimal feedback, offering a scalable and practical solution for 6G spatial multiple access in the upper mid-band FDD system.