Bridging Neural Networks and Wireless Systems with MIMO-OFDM Semantic Communications

TL;DR

This work investigates the gap between theory and practice in MIMO-OFDM semantic communications, focusing on PA nonlinearity and PAPR as sources of performance loss. Through a hardware prototype and end-to-end neural encoders/decoders, it shows that frequency-selective channels and per-symbol SNR variations largely drive deviations from simulated AWGN performance. It demonstrates that simple mitigation, such as symbol shuffling, can align real-world results with theory, and that PAPR-aware training yields tangible gains in nonlinear power regimes, thereby enabling near-theoretical performance in practical systems. The findings offer actionable guidance for channel-adaptive semantic designs and provide open-source code and hardware details to accelerate real-world deployment of semantic communications.

Abstract

Semantic communications aim to enhance transmission efficiency by jointly optimizing source coding, channel coding, and modulation. While prior research has demonstrated promising performance in simulations, real-world implementations often face significant challenges, including noise variability and nonlinear distortions, leading to performance gaps. This article investigates these challenges in a multiple-input multiple-output (MIMO) and orthogonal frequency division multiplexing (OFDM)-based semantic communication system, focusing on the practical impacts of power amplifier (PA) nonlinearity and peak-to-average power ratio (PAPR) variations. Our analysis identifies frequency selectivity of the actual channel as a critical factor in performance degradation and demonstrates that targeted mitigation strategies can enable semantic systems to approach theoretical performance. By addressing key limitations in existing designs, we provide actionable insights for advancing semantic communications in practical wireless environments. This work establishes a foundation for bridging the gap between theoretical models and real-world deployment, highlighting essential considerations for system design and optimization.

Figures (5)

• Figure 1: Basic block diagram comparing conventional digital communications and semantic communications system models. Conventional systems rely on separate source coding, channel coding, and modulation stages, whereas semantic communications adopt an end-to-end approach. This approach directly maps source data to wireless symbols, bypassing intermediate bitstreams, and reconstructs the data from noisy symbols. While it enables goal-oriented training, it is more sensitive to channel characteristics.
• Figure 2: System architecture of a semantic communications system with a MIMO-OFDM prototype setup. Red boxes indicate practical issues arising from the radio hardware and communication channels.
• Figure 3: Error plot from the wireless prototype, illustrating varying noise levels across subcarriers and data streams. Shuffling allocates symbols randomly across subcarriers, time slots, and streams to prevent concentration of poor channels into specific ranges of symbols. In the shuffled plot, the subcarrier index indicates where symbols would be allocated without shuffling.
• Figure 4: Constellation diagram from various PAPR-restricted semantic models with corresponding PSNR and PAPR values of the OFDM waveform with 72 subcarriers. Baseline, weak, and strong indicate the level of PAPR regulation of the semantic model, where the model is trained with PAPR loss weights of 0, 1/32768, and 1/4096, respectively.
• Figure 5: (a) Linear region result (MIMO-OFDM, Sim vs. OFDM-Semantic vs. OFDM-LDPC). (b) Nonlinear region result (Sim vs. low-PAPR vs. baseline models). In (b), the solid line and dotted line indicate the power-PSNR curve and the power-Rx SNR curve, respectively.