Uncertainty Decomposition for Machine-Learning-Based QoT Estimation in Coherent Optical Networks

Abstract

Machine learning (ML) has shown strong performance for Quality of Transmission (QoT) estimation in optical networks, but reliability under out-of-distribution (OOD) conditions remains a key operational concern. We propose an uncertainty-aware QoT framework that decomposes predictive uncertainty into aleatoric and epistemic components using a heteroscedastic neural predictor with a post-hoc Laplace approximation. The Laplace epistemic-uncertainty step can be fitted offline around a trained heteroscedastic predictor, while online inference remains deterministic. Across multiple OOD scenarios, the decomposition reveals complementary roles: aleatoric uncertainty mainly tracks modeled noise intensity, while epistemic uncertainty reflects model unfamiliarity under distribution shift. The empirical results indicate that uncertainty behavior differs clearly between familiar and unfamiliar regimes, which has direct implications for safe network operation. These support the conceptual two-regime operational policy: use aleatoric uncertainty for adaptive margining in trusted ML regions, and trigger fallback to a physics-based baseline when epistemic uncertainty is high. Overall, the results demonstrate the practical value of uncertainty decomposition for safer QoT estimation.