View article

Analytical models for quality of transmission (QoT) estimation require safety design margins to account for uncertain knowledge of input parameters. We propose and evaluate a design procedure that gradually decreases these margins in the presence of multiple physical-layer uncertainties (namely, connector loss, erbium-doped fiber amplifier gain ripple, and fiber type) by leveraging monitoring data to build a probabilistic machine-learning-based QoT regressor. We evaluate the savings from margin reduction in terms of occupied spectrum and number of installed transponders in the C and C+L bands and demonstrate that 4%–12% transponder/spectrum savings can be achieved in realistic network instances by simply leveraging the SNR monitored at receivers and paying off a low increment in the lightpath disruption probability (at most 1%–4%).