Attention-assisted autoencoder neural network for end-to-end optimization of multi-access fiber-terahertz communication systems

Abstract

We propose an end-to-end (E2E) fiber-terahertz (THz) integrated communication system based on an attention-assisted multi-access autoencoder (AMAE) neural network. The AMAE neural network comprises artificial neural networks (ANNs) that function as transmitters (T-ANNs), channel models, and receivers (R-ANNs) for multiple users. By connecting the computational graph of multiple T-ANNs and R-ANNs, we jointly optimize the AMAE to facilitate E2E multi-access communication. Attention mechanisms guide the optimization process to achieve fair and efficient power allocation and orthogonality among different users. We experimentally evaluated the performance of our proposed E2E framework in a 60 Gbit/s multi-channel (1, 5, and 10 km) fiber-THz hybrid system. The results indicate that our AMAE approach outperforms the conventional single-carrier quadrature amplitude modulation scheme by over 3 dB in receiver sensitivity and 11 Gbit/s in capacity under the 20% soft-decision forward error correction threshold in the same-channel back-to-back condition. Additionally, under the performance balance constraint, our approach achieves a transmission speed of 60 Gbit/s within a 10 GHz bandwidth in the multi-channel setting.