WDM System Stimulated Raman Scattering Spectrum and Tilt Prediction Using CNN-Based Transfer Learning

Advanced optical network technologies, proposed to fulfil the projected demands of future communication networks, are riddled with numerous challenges. Conventional methods for addressing these issues often face scalability limitations in disaggregated systems, prompting a growing reliance on data-driven approaches. Nevertheless, accurately modeling physical phenomena, such as stimulated Raman scattering (SRS), remains difficult since conventional physics-based models can falter due to variations in transmission spans. Machine learning-based models are promising in this regard, but they require huge datasets for enhanced versatility. Another alternative approach is transfer learning (TL), which minimizes the need for extensive datasets by utilizing pre-trained models that incorporate relevant domain knowledge. In this paper, we design and evaluate transfer learning models based on convolutional neural networks (CNN). We use these models to predict the Raman tilt spectra across four fiber configurations in two different testbeds. We achieve less than 0.162 dB mean absolute error which is largely limited by the measurement accuracy.