Input refinement with incremental learning for accurate digital twin-enabled self-driven QoT optimization in optical networks

Reliable quality of transmission (QoT) prediction is essential to ensure that services operate near their optimal working point, maximizing network capacity and efficiency while avoiding large design margins. A major source of QoT prediction error lies in the uncertain modeling of nonlinear fiber impairments, particularly the Kerr effect and stimulated Raman scattering, which are difficult to monitor with existing hardware. Accurate modeling of these effects requires knowledge of fiber insertion loss. Additionally, the gain spectrum of optical amplifiers, which governs amplified spontaneous emission noise, is not directly monitored. Further uncertainties stem from service end-to-end impairments, including wavelength selective switch (WSS) filtering-induced noise and from device-specific back-to-back (B2B) transponder performance, both of which are typically unknown at deployment time. To address these challenges, we propose incremental learning-based input refinement (IIR), a parameter estimation method that leverages multiple network snapshots to jointly refine fiber insertion losses, amplifier gain spectra, and end-to-end offset noise (comprising WSS filtering and transponder B2B contributions). IIR is applied incrementally during routine, live network (re-)optimization, requiring no intrusive measurements. The proposed method is validated through (1) simulations using experimental data on a C-band ring network and (2) experimental validation on a C-band mesh network. Results show that IIR significantly improves the estimation of key physical-layer parameters, enhancing QoT prediction accuracy and enabling effective closed-loop power optimization.