Qi Xu, Ran Gao, Xiaolong Pan, Xun Liu, Fei Wang, Huan Chang, Wei yan, yingyan zhang, Zhipei Li, Dong Guo, Lei Zhu, Qi Zhang, Yu Liu, Wei Cui, Xiangjun Xin
We propose and experimentally demonstrate an end-to-end (E2E) Tomlinson-Harashima precoding (THP) scheme for faster-than-Nyquist (FTN) coherent WDM transmission. A complex-valued conditional generative adversarial network (CV-CGAN) trained on measured complex waveforms is used as a differentiable channel emulator, enabling joint optimization of the transmitter feedback equalizer (FBE) and receiver feed-forward equalizer (FFE). The proposed scheme is experimentally validated in a 30-channel FTN-WDM system over 80-km single-mode fiber (SMF). The experimental results demonstrate that the proposed CV-CGAN-based E2E THP scheme achieves the same bit error rate (BER) performance with optical signal-to-noise ratio (OSNR) sensitivity improvements of 0.5 dB and 1.5 dB compared to conditional generative adversarial network (CGAN)-based E2E THP and traditional minimum mean-square-error (MMSE)-based THP scheme, respectively, under a 20% soft-decision forward error correction (SD-FEC) threshold.