ML-Driven QoT Estimation for Adaptive P2MP Optical Network-as-a-Service in 6G X-Haul

Point-to-Multipoint (P2MP) coherent transmission enables efficient converged metro-access networks for future B5G/6G services, but dynamic routing, Distributed Units (DUs) placement, and spectrum reconfiguration introduce physical- layer impairments that make conventional analytical Quality- of-Transport (QoT) estimation inaccurate and computationally expensive. This work proposes a scalable and topology-agnostic QoT prediction framework that formulates BER estimation as a regression problem using physical-layer and network related fea- tures. Models are trained on a small topology and directly applied to larger unseen networks to evaluate cross-topology generaliza- tion under limited training data and noisy telemetry, reflecting realistic operational conditions. Results show that lightweight boosting-based regression, particularly XGBoost, achieves near perfect BER prediction with sub-millisecond inference, enabling real-time feasibility checks without executing full physical-layer simulations. These findings demonstrate that ML-driven QoT regression can replace analytical evaluation and serve as a practical enabler for intelligent and adaptive Optical Network- as-a-Service (ONaaS) for future 6G transport