Neural Codebook Design for Network Beam Management

TL;DR

This work tackles the CSI and beam management bottlenecks in large hybrid MIMO/MU-MIMO systems by introducing Network Beamspace Learning (NBL), an end-to-end neural framework that designs SSB and CSI-RS codebooks to mitigate inter-cell interference and maximize network spectral efficiency. By representing beamformers in beamspace and backpropagating through the entire beam management chain, NBL learns codebooks that support full-rank MIMO and interference-aware operation with limited cross-cell signaling. Experiments on raytraced urban scenarios show >10 dB gains in beam alignment and >25% improvements in network spectral efficiency, with robust generalization across array geometries, deployment sites, and even out-of-distribution environments. The results highlight the potential of integrating AI-driven, end-to-end codebook design into 6G X-MIMO deployments to reduce training overhead and enhance network throughput.

Abstract

Obtaining accurate and timely channel state information (CSI) is a fundamental challenge for large antenna systems. Mobile systems like 5G use a beam management framework that joins the initial access, beamforming, CSI acquisition, and data transmission. The design of codebooks for these stages, however, is challenging due to their interrelationships, varying array sizes, and site-specific channel and user distributions. Furthermore, beam management is often focused on single-sector operations while ignoring the overarching network- and system-level optimization. In this paper, we proposed an end-to-end learned codebook design algorithm, network beamspace learning (NBL), that captures and optimizes codebooks to mitigate interference while maximizing the achievable performance with extremely large hybrid arrays. The proposed algorithm requires limited shared information yet designs codebooks that outperform traditional codebooks by over 10dB in beam alignment and achieve more than 25% improvements in network spectral efficiency.

Figures (13)

• Figure 1: A block diagram of the system-level model and $\text{NBL}$ codebook roles. The proposed solution modifies the green blocks, while the orange and blue blocks refer to assumed and standard operations respectively. The end-to-end learning framework is shown by the green dashed box that encapsulates the codebook operations. The assumed operations outside of the end-to-end learning are not critical but allow for system-level evaluation.
• Figure 2: Users are allocated to the cells during SSB transmission. During CSI-RS transmission, only interference during a UE's selected CSI-RS causes interference. Therefore, codebooks should be designed so that the first green CSI-RS does not interfere strongly with the orange UE and vice versa.
• Figure 3: A visual representation of the $\text{NBL}$ inference and training within the network. SSB feedback is shared between cells to jointly generate SSB and CSI-RS codebooks. Gradients are backpropagated through the entire beam management system to maximize the achievable rate during CSI-RS.
• Figure 4: A comparison of the powermaps in two environments with different base station heights and carrier frequencies. Scene 1 corresponds to Munich, Germany operating at $10$GHz and base stations mounted at $40$m tall and $450$m intersite distance. The OOD environment in scene 2 is Paris, France operating at $20$ GHz, with $27$m high serving cells and $370$m intersite distance. A low resolution power map is rendered on each scene in the overlapping region from the perspective of base station $1$ in each of the scenes.
• Figure 5: The CDF of the reported RSRP using $\text{NBL}$ and DFT codebooks of different sizes $L_{\text{max}}$. The broadcast SSB refers to a single, non-beamformed signal and is shown for comparison. The $\text{NBL}$ codebooks result in an average of $>5$dB gain over DFT codebooks and reduced user outage rates.