Simulation-to-Experiment Aided Transfer Learning in Data Driven Nonlinear Equalization for Raman-Amplified Coherent Systems

<jats:p>This study evaluates an LSTM-based nonlinear equalizer trained with only about 14,000 experimental symbols from backward-pumped Raman-amplified, single-wavelength 54.5-GBaud 16-QAM links. Training solely on the limited measurements provides only modest BER gains over conventional digital back-propagation (DBP). To address data scarcity, a Raman-specific modified nonlinear Schrödinger equation generates synthetic training traces. The LSTM is pretrained on a unified simulated dataset, then adapted to experiments through transfer learning. This simulation-to-experiment approach matches or surpasses optimized Raman-aware DBP, reducing BER by up to twofold over 3200- and 4000-km links, and highlighting the promise of digital-twin-assisted equalization for distributed Raman transmission systems in practical deployments.</jats:p>