Machine Learning Model Training Framework for Nonlinear Signal-to-Noise Ratio Estimation in Heterogeneous Optical Networks

A computationally efficient framework is presented for calculating features used to train machine learning (ML) models for estimating the nonlinear signal-to-noise ratio (
$S\!N\!R_{N\!L}$
) in heterogeneous optical networks. Data used for training and testing is obtained from a first-order perturbation analytical model, which is calculated for over 500,000 distinct system configurations, covering a wide range of system and channel compositions. System configurations vary based on span lengths, number of spans, number of wavelength division multiplexed (WDM) channels, channel spacings, modulation formats, shaping rates, and symbol rates. Five ML models have been trained using features extracted from the nonlinear phase noise generated by signal-signal interaction between WDM channels. Model comparison suggests that ensembles of regression trees produce highest estimation accuracy. A robust model building method is presented that aggregates important features from three ensemble models with boosting and shows universal application to all considered training cases. The impact from the extent of heterogeneity and training diversity in terms of number of WDM channels, symbol rate, and total distance covered in training is explored. Estimation results demonstrate the benefit of considering heterogeneous system configurations in training and indicate high accuracy and generalization potential to arbitrary system configurations.