OSNR/GSNR Prediction in Brownfield Links via a DLM-Anchored Hybrid Physics/ML Model

Brownfield optical links present unique challenges for performance prediction due to legacy infrastructure, unknown fiber parameters, and heterogeneous equipment. Accurate prediction of per-channel power, optical signal-to-noise ratio (OSNR), and generalized signal-to-noise ratio (GSNR) is critical for efficient network planning and service provisioning. Traditional physics-based models often fail in brownfield environments because they rely on precise knowledge of span losses, amplifier gains, and inline amplifier (ILA) locations, which are frequently unavailable or inaccurate. Machine learning (ML) approaches can learn from data but may lack physical interpretability and generalize poorly to unseen configurations. This work introduces a hybrid physics/ML framework anchored by a digital line model (DLM) to bridge this gap. The DLM provides a physics-based reference that calibrates span and ILA boundaries, enabling the ML component to correct residual errors. The framework is evaluated on brownfield links for single-channel and optical spectrum as a service (OSaaS) provisioning, demonstrating OSNR and GSNR prediction errors of no more than 0.39 dB and 0.43 dB, respectively. This digest outlines the methodology, results, and implications of this approach.

The proposed framework combines a physics-based DLM with a machine learning correction stage. The DLM simulates signal propagation through the optical link using standard fiber and amplifier models, but its parameters (e.g., span losses, ILA positions) are calibrated using measured data. Calibration is performed by minimizing the discrepancy between DLM predictions and a set of reference measurements, which may include per-channel power or OSNR at specific points. The calibrated DLM then serves as a feature extractor and baseline predictor. An ML model, such as a neural network or gradient boosting regressor, is trained on the residuals between DLM predictions and ground truth to capture unmodeled effects like nonlinearities, filter penalties, or equipment-specific behaviors. The hybrid model outputs per-channel power, OSNR, and GSNR. The framework is designed for brownfield links where full physical knowledge is lacking. The calibration process is formulated as an optimization problem: given measured data $\mathbf{y}$ and DLM parameters $\mathbf{\theta }$, find

, where $f_{\text{DLM}}$ is the DLM function. The ML correction is then , with $\mathbf{x}$ being features from the DLM and measurements. The overall workflow is illustrated in the Mermaid diagram below.