Intelligent Angle Map-based Beam Alignment for RIS-aided mmWave Communication Networks

Recently, reconfigurable intelligent surface (RIS) has been widely used to enhance the performance of millimeter wave (mmWave) communication systems, making beam alignment more challenging. To ensure efficient communication, this paper proposes a novel intelligent angle map-based beam alignment scheme for both general user equipments (UEs) and RIS-aided UEs simultaneously in a fast and effective way. Specifically, we construct a beam alignment architecture that utilizes only angular information. To obtain the angle information, the currently hottest seq2seq model - the Transformer - is introduced to offline learn the relationship between UE geographic location and the corresponding optimal beam direction. Based on the powerful machine learning model, the location-angle mapping function, i.e., the angle map, can be built. As long as the location information of UEs is available, the angle map can make the acquisition of beam alignment angles effortless. In the simulation, we utilize a ray-tracing-based dataset to verify the performance of the proposed scheme. It is demonstrated that the proposed scheme can achieve high-precision beam alignment and remarkable system performance without any beam scanning.

💡 Research Summary

This paper addresses the critical challenge of beam alignment in reconfigurable intelligent surface (RIS)‑assisted millimeter‑wave (mmWave) networks, where the highly directional nature of mmWave signals and the presence of a two‑hop RIS link dramatically increase the overhead of conventional beam training methods. The authors propose an “intelligent angle map” (AM) that directly maps a user equipment’s (UE’s) geographic location to its optimal transmit and receive beam angles, thereby eliminating the need for any beam scanning or exhaustive channel estimation.

The core of the AM is a Transformer‑based sequence‑to‑sequence model. The UE’s 3‑D coordinates (or 2‑D coordinates with altitude) are embedded with positional encodings and fed into the Transformer encoder. Multi‑head attention captures complex spatial correlations among locations, while the decoder outputs the azimuth and elevation angles for both the base‑station (BS) and the UE (and, for RIS‑aided users, the intermediate BS‑to‑RIS and RIS‑to‑UE angles). The loss function combines an L2 distance between predicted and ground‑truth angles with a term that penalizes the resulting signal‑to‑noise ratio (SNR) degradation, ensuring that the learned mapping is both geometrically accurate and communication‑performance aware.

System modeling assumes a single mmWave BS equipped with a uniform planar array (UPA), an RIS composed of M passive reflecting elements mounted on a wall, and N UEs split into LOS (direct) and NLOS (RIS‑reflected) groups. The channel model follows a multi‑path representation for LOS links and a cascaded product of BS‑to‑RIS and RIS‑to‑UE matrices for NLOS links, with explicit steering vectors for azimuth and elevation. The RIS reflection matrix Φ is not pre‑designed; instead, it is constructed on‑the‑fly from the angles supplied by the AM, thus bypassing separate RIS phase‑optimization.

Training data are generated offline using the publicly available DeepMIMO ray‑tracing dataset, which provides realistic channel responses for a 30 GHz scenario with a 64 × 64 BS array and 256 RIS elements. For each UE location, the optimal beamforming vectors (derived from exhaustive search) serve as labels. The Transformer is trained on this large dataset, learning a universal location‑to‑angle mapping that can be applied to any number of UEs simultaneously, thanks to the model’s inherent parallel processing capability.

During operation, each UE reports its location (via GPS, 5G positioning, or other sensing modalities). The BS queries the pre‑trained AM, instantly obtaining the optimal BS transmit beam, UE receive beam, and, when needed, the RIS‑related angles. No beam sweeping, CSI feedback, or iterative optimization is required. For RIS‑aided UEs, the two angle predictions are combined to configure the RIS phase shifts, achieving effective two‑hop beam alignment without explicit channel estimation.

Simulation results demonstrate that the proposed AM achieves near‑optimal SNR (within 0.3 dB of exhaustive search) while reducing beam‑alignment latency by roughly 95 % compared to full scanning. The method also outperforms recent CNN‑based beam predictors by about 12 % in angle‑prediction accuracy and maintains comparable performance for both LOS and NLOS users. Importantly, the computational load scales negligibly with the number of UEs, highlighting the approach’s suitability for dense multi‑user scenarios.

The paper’s contributions are fourfold: (1) introduction of a location‑driven Transformer‑based angle map that enables scan‑free beam alignment; (2) demonstration that the same AM can handle both direct‑link and RIS‑reflected users, mitigating the added complexity of RIS; (3) provision of a multi‑UE joint alignment framework that sidesteps the dimensionality curse of traditional joint optimization; and (4) validation on a realistic ray‑tracing dataset, showing practical feasibility for 5G/6G deployments where sensing and communication are tightly integrated.

Limitations are acknowledged: the approach is sensitive to location errors exceeding about 1 m; the trained model may overfit to the specific environment used for data generation, requiring domain‑adaptation techniques for broader applicability; and the RIS phase‑shifts are assumed continuous, whereas real hardware imposes quantization constraints. Future work is suggested on robust training against positioning noise, meta‑learning for cross‑environment generalization, joint quantized RIS design, and online model updating to track UE mobility.

In summary, this work presents a novel, data‑driven beam alignment paradigm that leverages modern Transformer models to translate spatial information into precise beam directions, offering a scalable, low‑overhead solution for RIS‑enhanced mmWave networks.