TU‐AcqNet: A Transformer‐U‐Net Framework for Robust Channel Acquisition in THz UM‐MIMO Systems

ABSTRACT Accurate channel acquisition remains a fundamental challenge in terahertz (THz) ultra‐massive multiple‐input multiple‐output (UM‐MIMO) communication systems, primarily due to near‐field propagation effects and the large‐scale nature of antenna arrays. To address the limitations of existing compressed sensing and deep learning‐based approaches, a novel framework referred to as transformer‐U‐Net acquisition network (TU‐AcqNet) is proposed. This hybrid model integrates multi‐head self‐attention‐based transformer encoders with a U‐Net‐inspired decoder to jointly capture global pilot signal dependencies and reconstruct high‐resolution estimates of channel parameters The novelty lies in its dual‐stage architecture that combines temporal sequence modelling with spatial feature reconstruction, enabling highly accurate channel parameter estimation even under sparse pilot constraints. TU‐AcqNet emphasizes dominant pilot features through attention mechanisms and estimates key channel parameters—including angles of arrival, path distances, and complex path gains—with high resolution. The proposed scheme achieves a normalized mean square error (NMSE) improvement of up to 6 dB and increases spectral efficiency by more than 4 bits/s/Hz in high signal‐to‐noise ratio scenarios. Overall, the framework yields up to 80% reduction in NMSE relative to state‐of‐the‐art baselines, highlighting its potential for practical deployment in next‐generation THz UM‐MIMO systems.