AI-Enabled Semantic Communication in the Terahertz Band

Abstract

Semantic communication, which focuses on transmitting meaning rather than ex act bit sequences, has emerged as a promising paradigm for next-generation wire less systems. In parallel, terahertz (THz) communication has gained significant attention as a key enabler of ultra-high data rates due to its abundant bandwidth. However, deploying semantic communication in THz bands introduces significant challenges, including severe propagation effects, costly channel state information (CSI) acquisition, and strong frequency selectivity. This thesis develops a CSI-free semantic communication framework for THz systems, spanning generalization abil ity, physical-layer equalization, and resource allocation. The thesis makes three main contributions. First, we show that deep learning-enabled semantic commu nication (DeepSC) models trained under additive white Gaussian noise (AWGN) can generalize to complex THz fading environments without retraining, with perfor mance largely governed by the effective noise distribution rather than the channel structure. Second, using a Bayesian formulation, we show that the structured de pendencies inherent in semantic representations render conventional symbol-wise equalization suboptimal. We propose a lightweight sequence-aware neural equalizer (NE) that leverages inter-token correlations via transformer self-attention, oper ating without CSI across single-input single-output (SISO), single-input multiple output (SIMO), and multiple-input multiple-output (MIMO) configurations. The proposed approach consistently outperforms conventional minimum mean square er ror (MMSE) equalization, under both perfect and imperfect CSI. Third, we address semantic-aware sub-band allocation in frequency-selective THz channels, where op timal assignment depends jointly on sentence importance and channel quality. We formulate an importance-weighted optimization problem capturing these factors and solve it using imitation learning, achieving near-optimal performance with signifi cantly reduced complexity. Collectively, the proposed framework demonstrates the feasibility of CSI-free semantic communication over THz channels and highlights the potential of data-driven semantic approaches for next-generation wireless systems.