Gray-Box Modeling for Predictive Gain and Loss in Optical Line Systems

Optical fiber networks have become the backbone of modern telecommunications due to their high capacity, low latency, and cost-efficiency. Ensuring reliable Quality of Trans- mission (QoT) in these networks requires accurate estimation of performance metrics prior to lightpath deployment. Traditional approaches often rely on worst-case assumptions to account for uncertainties that might arise, for instance, from amplified spontaneous emission (ASE) noise introduced during optical amplification, leading to conservative de- sign margins that reduce spectral efficiency and increase operational costs. This thesis investigates the use of gray-box modeling techniques, which integrate analytical models with Machine Learning (ML), to enhance the prediction accuracy of gain and loss profiles across an Optical Line System (OLS). The work evaluates multiple model architectures to estimate amplifier gain profiles and fiber span losses, using a finite set of measured data to train and validate the models. Comparative analysis against benchmark models demonstrates that gray-box approaches can significantly reduce prediction error, enabling more accurate QoT estimation and more efficient resource allocation.