Abstract
The transition to 6G communications necessitates the use of the Terahertz (THz) spectrum to achieve ultra-high data rates. However, the extreme pathloss at these frequencies requires massive antenna arrays and precise beamforming, which traditionally incurs significant pilot overhead and latency. This paper presents a deep learning framework that utilizes sub-6GHz channel state information as a proxy to estimate THz channel parameters and predict optimal beamforming vectors. By exploiting the spatial correlation between these frequency bands, the proposed method achieves near-optimal spectral efficiency with minimal overhead.
Introduction
As 6G targets data rates in the Gb/s and Tb/s range, the Terahertz spectrum (0.1–10 THz) becomes indispensable. The primary challenge in THz communication is the requirement for highly directional "pencil beams" to overcome severe atmospheric attenuation and pathloss. Traditional channel estimation techniques, such as Least Squares (LS) or Linear Minimum Mean Square Error (LMMSE), are computationally expensive and require a large number of pilot signals, leading to high latency that is incompatible with real-time applications like AR/VR.
Proposed Framework
The authors propose a dual-band system where the base station and user equipment operate on both sub-6GHz and THz frequencies. Since both signals propagate through the same physical environment, their channel characteristics—such as angles of arrival/departure and delays—are inherently correlated.
The framework employs a two-stage deep learning architecture:
- THz Channel Factor Estimator (CNN): A Convolutional Neural Network processes the sub-6GHz channel matrix (decomposed into magnitude and phase) to estimate seven critical THz channel factors: pathloss, time of arrival (ToA), phase change, and azimuth/elevation angles of arrival and departure (AoA/AoD).
- Beamformer Predictor (Dense NN): A subsequent dense neural network uses these estimated factors to directly select the optimal beamformer index from a predefined codebook.
Experimental Setup
The model was validated using synthetic data generated via MATLAB’s Raytracing Communications Toolbox, simulating a realistic outdoor environment in downtown Chicago. The dataset comprised 100,000 user positions, with 80,000 used for training and 20,000 for testing. The simulation incorporated atmospheric effects like cloud and fog attenuation to reflect real-world THz propagation challenges.
Key Results
The CNN-based estimator achieved high precision, with mean errors for angular parameters under 6 degrees and an exceptionally low error for Time of Arrival ($3.8 \times 10^{-18}$ seconds). Notably, the sub-6GHz-based approach outperformed baseline models using THz uplink data, as the latter often suffered from excessive noise due to high attenuation.
In terms of system performance, the "Top-3" beam prediction reached spectral efficiency levels nearly identical to the theoretical Extensive Search Upper Bound (UB), which assumes perfect channel knowledge.
Conclusion and Future Work
This research demonstrates that deep learning can effectively bridge the gap between low-frequency channel data and high-frequency beamforming requirements. While the current study focuses on static environments, future work will address high-speed mobility and hardware-specific impairments such as phase noise. The ability to infer THz parameters from sub-6GHz signals opens new avenues for proactive blockage prediction and low-latency 6G networking.
Summary updated