Semantic Communications in the THz Band

Technical Summary: Semantic Communications in the THz Band

Problem Statement
Future wireless networks require the Terahertz (THz) band to achieve ultra-high data rates and massive connectivity. However, THz channels suffer from severe path loss, frequency selectivity, molecular absorption, and rapid fading, making accurate Channel State Information (CSI) acquisition difficult and costly, particularly at low Signal-to-Noise Ratios (SNRs). While Deep Learning-based Semantic Communication (DeepSC) has demonstrated robustness in Additive White Gaussian Noise (AWGN) and simplified fading environments, its performance in THz conditions remains largely unexplored. Existing learning-based detectors often rely on pilots or explicit CSI, and traditional equalizers like Minimum Mean Square Error (MMSE) require accurate CSI, which is unreliable under mobility. Furthermore, current semantic systems often require retraining when channel distributions change, creating a barrier to deployment in dynamic THz environments.

Methodology
The authors propose a fully data-driven, CSI-free framework for text-based semantic communication over THz channels. The system integrates a DeepSC architecture with a proposed lightweight Neural Detector (ND).

• System Architecture: The framework transmits the semantic meaning of a sentence using a semantic encoder and a channel encoder. The transmitted symbols pass through a realistic THz MIMO channel modeled using the TeraMIMO simulator, which incorporates Saleh–Valenzuela clustered multipath, frequency-dependent molecular absorption (HITRAN data), and subarray beamforming.
• Neural Detector (ND): To eliminate the need for pilots and CSI estimation, the authors replace conventional equalization with a lightweight Transformer-based ND placed before the channel decoder. The ND processes the received sequence (reshaped to capture real and imaginary parts) using 1D convolutions to capture local distortions, followed by fixed sinusoidal positional encoding and two Transformer encoder layers with self-attention. This allows the model to exploit global dependencies between words in a sentence rather than denoising symbols independently. The detector has approximately 100,802 trainable parameters.
• Training Strategy: The system employs a two-stage training procedure:
  1. AWGN Pretraining: The DeepSC encoder-decoder is trained end-to-end over an AWGN channel to optimize semantic transmission.
  2. ND Training: The DeepSC components are frozen. The ND is trained over THz channels to map corrupted observations directly to the clean transmitted channel-symbol sequence, minimizing Mean Squared Error (MSE) against the AWGN-trained target. This approach relies on the insight that, under effective equalization, the fading channel can be viewed as an additive Gaussian perturbation, allowing an AWGN-trained decoder to function effectively.

Key Contributions

1. Generalization of AWGN-trained DeepSC: The authors demonstrate that DeepSC models trained solely under AWGN generalize effectively to THz block- and fast-fading channels when receiver-side compensation is applied, eliminating the need for channel-specific retraining.
2. Pilot-Free Neural Detector: A lightweight, Transformer-based ND is proposed that recovers continuous channel-symbol sequences without requiring pilots, explicit CSI, or matrix inversions.
3. Robustness Analysis: The study analyzes the system's performance under imperfect CSI and frequency offsets, showing that the proposed framework outperforms traditional coded systems and MMSE equalization in these challenging conditions.

Results
Simulations were conducted at 0.3 THz over a 10 GHz bandwidth using an indoor THz scenario with 4096×4096 antenna elements.

• Performance vs. Traditional Systems: At 0.3 THz, the DeepSC system with the proposed ND outperforms a throughput-matched traditional coded communication baseline (LDPC-coded 64-QAM with MMSE) across 0–12 dB SNR. The improvement is significant, achieving more than 50 percentage-points higher Bilingual Evaluation Understudy unigram (BLEU-1) score.
• ND vs. MMSE: The proposed pilot-free ND outperforms MMSE equalization using both perfect CSI and pilot-based (imperfect) CSI. It also demonstrates superior robustness to frequency offsets up to ±50 MHz, whereas MMSE performance degrades significantly when the CSI does not match the operating frequency.
• Generalization: The AWGN-trained model maintains high performance across different fading update rates (block, sentence, and word-level fading) and different noise distributions (Gaussian, Laplace, Student-t), confirming that semantic performance is more sensitive to post-equalization noise statistics than to the specific fading profile.
• Latency: In terms of processing time on an NVIDIA V100 GPU, the proposed ND is approximately 1% faster than LS-based MMSE equalization and requires no explicit CSI estimation overhead.

Significance
The paper claims that this work addresses a critical gap in THz communications by enabling fully data-driven semantic transmission without the costly and unreliable overhead of pilot signals or CSI acquisition. By showing that AWGN-trained semantic models can operate effectively over complex THz channels with a lightweight, pilot-free receiver, the authors highlight the resilience of semantic communication to THz-specific impairments. This approach offers a viable path toward robust, high-rate THz links where traditional channel estimation methods fail, particularly in dynamic mobility scenarios.