Deep Graph Learning for QoT Estimation of Unseen Optical Sub-Network States: Capturing the Crosstalk Impact on the In-Service Lightpaths

In this work, deep graph convolutional neural networks (DGCNN) are applied for estimating the quality-of-transmission (QoT) of unseen network states in elastic optical networks (EONs) in the presence of physical layer impairments (PLIs), including inter- and intra-channel crosstalk (XT). The objective is to find a DGCNN-QoT model that accurately estimates network state feasibility. A network state is considered feasible if the QoT of the in-service lightpaths and of the lightpath under provisioning is sufficient; that is, the DGCNN does not only infer about the feasibility of an unestablished lightpath but also whether the feasibility of the in-service lightpaths will be affected by the establishment of a new lightpath due to XT. As DGCNN model generalization over unseen graphs is known to be negatively affected by the number of possible graphs and their dimensionality, problem uncertainty and complexity is reduced by formulating the QoT estimation problem over sub-network states, capturing only the spatio-temporal correlations that are relevant to the unestablished lightpath at decision time. DGCNN model accuracy is compared to a state-of-the-art deep neural network (DNN) model trained only over per-lightpath information. It is shown that DGCNN achieves accuracies above 92%, while DNN performs poorly with accuracies as low as 77%, as it fails to infer about the feasibility of in-service connections; an indicator of the importance of explicitly considering during the QoT model training, not only the lightpath patterns, but also the network-state patterns capturing the XT effect. Importantly, it is demonstrated that deep graph learning is a promising approach towards accomplishing this objective.