Machine learning model based on the time domain regular perturbation-based theory for performance estimation in arbitrary heterogeneous optical links

To meet the required high demands for the capacity of optical networks, there are several efforts in recent years to further reduce the system margin. To achieve this goal, a fast and reliable QoT estimation tool is needed. The key module of such QoT tool is the nonlinear interference variance estimation. This paper presents a novel machine learning tool to speed up the exact model by six orders of magnitude without scarifying the accuracy and the scalability of this semi-analytical model. Moreover, to further enhance the scalability of the model, we used the cross-correlation functions to re-write the equations. The proposed ML-based framework KerrNet, based on a bank of small ANNs, can handle any arbitrary heterogeneous link up to ten thousand km composed of different fiber span. The transmitting C-band WDM channels in both fully loaded and sparsely occupied configurations are evaluated. The crucial steps for the machine learning algorithm to converge, which are the data preparation and the choice of training data, are presented in detail.