Answers are grounded in this paper and cite their sources.
Multipath interference (MPI) is a critical impairment limiting the performance in short-reach intensity modulation and direct detection (IM/DD) systems, especially for multi-level pulse amplitude modulation (PAM) formats. In data center scenarios, MPI is further exacerbated by connector aging, end-face contamination, and the widespread use of cost-effective lasers in high-density fiber links, meanwhile its severity varies significantly across different links and evolves dynamically over time. In this paper, we proposed a synergistic feedforward neural network (FNN) and fluctuation smoothing (FS) scheme for MPI mitigation, termed Synergistic FNN-FS. To improve robustness against varying interference conditions, the FNN is trained using a mixed signal-to-interference ratio (SIR) dataset, which alleviates overfitting to specific MPI severity levels. The trained FNN performed initial equalization on MPI-impaired signals, and then the fluctuation smoothing module suppressed residual fluctuations of FNN equalized signals caused by the time-varying stochastic characteristic of MPI and the generalization compromise introduced by the mixed-SIR dataset. The proposed scheme exhibited functional complementarity of the FNN and FS modules, effectively mitigating MPI-induced impairments. Simulation results in a 28-GBaud IMDD PAM-4 system demonstrate the effectiveness of the proposed scheme. Under back-to-back (BTB) conditions with a 10-MHz laser linewidth, the proposed scheme achieves SIR tolerance improvements of approximately 1 dB and 3 dB compared with standalone FNN equalization and conventional RF-M, respectively, at the 7% HD-FEC threshold. In the 2-km standard single-mode fiber (SSMF) transmission scenario, the proposed scheme also maintains superior BER performance, providing receiver sensitivity gains of approximately 2 dB and 6 dB over FNN and CNN equalization, respectively, at a fixed SIR of 20 dB.
No takes yet. Share an insight, caveat, or question.
Huang et al. (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: