Graph Neural Network Based Hybrid Beamforming Design in Wideband Terahertz MIMO-OFDM Systems

TL;DR

The paper tackles beam squint in wideband Terahertz MIMO-OFDM by replacing costly true-time-delay solutions with a Graph Neural Network that jointly learns analog and digital beamformers. It introduces a bipartite graph representation with an analog node and K digital nodes, leveraging per-subcarrier channel edge features to perform message passing and produce beamformers that satisfy hardware constraints. Offline training targets maximizing spectral efficiency, while online inference scales with the number of subcarriers, enabling real-time adaptation. Experimental results show the GNN closely matches AMO and approaches fully digital performance while delivering significantly lower runtime and memory usage, and demonstrates strong resilience to beam squint as bandwidth grows. This method offers a practical, scalable alternative for hybrid beamforming in future high-frequency wireless systems.

Abstract

6G wireless technology is projected to adopt higher and wider frequency bands, enabled by highly directional beamforming. However, the vast bandwidths available also make the impact of beam squint in massive multiple input and multiple output (MIMO) systems non-negligible. Traditional approaches such as adding a true-time-delay line (TTD) on each antenna are costly due to the massive antenna arrays required. This paper puts forth a signal processing alternative, specifically adapted to the multicarrier structure of OFDM systems, through an innovative application of Graph Neural Networks (GNNs) to optimize hybrid beamforming. By integrating two types of graph nodes to represent the analog and the digital beamforming matrices efficiently, our approach not only reduces the computational and memory burdens but also achieves high spectral efficiency performance, approaching that of all digital beamforming. The GNN runtime and memory requirement are at a fraction of the processing time and resource consumption of traditional signal processing methods, hence enabling real-time adaptation of hybrid beamforming. Furthermore, the proposed GNN exhibits strong resiliency to beam squinting, achieving almost constant spectral efficiency even as the system bandwidth increases at higher carrier frequencies.

Figures (7)

• Figure 1: Block diagram of a single-user MIMO-OFDM system with hybrid beamforming architecture at the BS and the UE.
• Figure 2: Graph representation of a hybrid beamforming structure with two types of nodes representing the analog and the digital beamformers, where each edge has an edge feature $\mathbf{e}_k$. Here we implement a message-passing mechanism where messages are denoted as $\mathbf{m}_{1,k}^a$ and $\mathbf{m}_{k,1}^d$.
• Figure 3: The overall architecture of the proposed GNN consists of $L$ updating layers, each integrating a message-passing layer and a node representation updating layer, followed by a final beamformer reconstruction layer.
• Figure 4: Convergence of the proposed GNN training for SNR = 0dB, $N_t=64$, $N_{\text{RF}}^t=4$, $K=8$, $f_c=300$GHz, and $B=30$GHz.
• Figure 5: Spectral efficiency achieved by different beamforming design algorithms with $K=64$ subcarriers, averaged over $10^3$ channel realizations.