ML-Based Physics-Guided Method for Low-Margin Nonlinear Interference Estimation in Optical Links

Modern-day Optical Transport Networks run over a heterogeneous fiber infrastructure, with increasingly tight design margins and dynamic channel loading conditions. As a result, the estimation of Nonlinear Interference (NLI) is a central requirement for performance prediction and automation. In dispersion-managed (DM) links, NLI is typically described as the sum of Self-Channel Interference (SCI) and Cross- Channel Interference (XCI), but traditional Gaussian Noise (GN) predictions become inaccurate. This research focuses on the semi-analytical NLI model recently introduced for SCI and XCI in DM systems, where the GN contribution is corrected through two key parameters: the spatial correlation factor Cij , accounting for nonlinear interactions between spans i and j, and the GN conservativity correction ρi, correcting the intrinsic overestimation of the GN model. As these terms lack closed- form expressions, we tested three regression Machine Learning models and proposed the PhysicsXGB model for predicting Cij and ρi separately for SCI and XCI. The analysis investigated the impact of physics-derived engineered features and span-level heterogeneity on regression accuracy, evaluating all models under realistic DM link configurations. The results showed that the proposed model enabled real-time and spectrally disaggregated NLI estimation with significant accuracy, providing a viable path towards a fast and reliable Digital Twin engine for optical networks.