Physics-Based Modeling for Hybrid Data-Driven Models to Estimate SNR in WDM Systems

Recently several machine learning methods have been proposed to estimate the SNR, based on launch data and other system factors. These data-driven methods typically require a large number of datasets for training and generally are not interpretable. In this paper, we propose an alternative approach that requires less data and is interpretable, specifically a hybrid algorithm combining a physical model with Gaussian process regression. We develop a measurement-informed physical model, systematically reducing the number of independent parameters based on the underpinning physics and improve the overall performance of the physical model marginally. The model is validated using measurements performed on a 15-channel wavelength-division multiplexed system propagating over 1,000 km of standard single-mode fiber. The proposed hybrid model is not only interpretable but also obtains better agreement with measurements than a Gaussian process regression model and a simple neural network model for a given number of training datapoints.