Learning-Based Hybrid Neural Receiver for 6G-V2X Communications

Neural receivers are emerging as key enablers for intelligent physical-layer design in next-generation wireless communications. However, the integration of learning-based architectures into practical communication systems requires careful consideration of standard compatibility and compu tational constraints. In this work, we propose a Hybrid Neural Receiver (H-NR) that operates on the frequency-domain OFDM resource grid and replaces the post-FFT baseband receiver chain, including resource grid demapping, channel estimation, equalization, demodulation, and channel decoding, within a unified end-to-end architecture. The proposed design integrates Transformer encoder blocks for global symbol inference with a Graph Neural Network (GNN) for structure-aware channel decoding. Extensive evaluations under varying vehicular speeds (0–120 km/h), carrier frequencies (5.9 and 28 GHz), and standardized UMa and CDL channel models demonstrate consistent gains of approximately 0.5 dB over the state-of-the-art TransRx neural receiver, particularly in scenarios where conventional receivers exhibit significant performance degradation. The proposed H-NR demonstrates efficient GPU utilization, achieving a throughput of 33.95 GFLOPs/s and a 121.7% improvement over the state-of-the-art TransRx. This performance, together with its ability to process diverse data modalities such as data, image, and audio, highlights its suitability for next-generation C-V2X systems