Simsek, Aybike2026-04-252026-04-2520252169-3536https://doi.org/10.1109/ACCESS.2025.3647668https://hdl.handle.net/11486/8442Evaluating the connectivity resilience of real-world networks through attack simulations is a time-consuming process. This study proposes a method that trains Graph Convolutional Networks (GCNs) on synthetic graphs-generated from real graphs-to predict the resilience curves (Normalized Largest Connected Component, NLCC) of real networks. The method was tested on six real-world interaction networks under degree-based and random attack strategies. Results show that GCN achieves Mean Absolute Error (MAE) values between 0.03 and 0.23, with higher accuracy under random attacks. Additionally, the entire process-including synthetic graph generation, attack simulations, and GCN training-takes approximately five times less time than direct simulations on real graphs with similar size and density. These findings demonstrate that the proposed approach offers both accurate and efficient resilience estimation.eninfo:eu-repo/semantics/openAccessResilienceRobustnessComplex networksControllabilityMeasurementAccuracyBiological system modelingPredictive modelsGraph convolutional networksAdaptation modelsgraph neural networksnetwork resiliencerobustness predictionArticle1321706121707210.1109/ACCESS.2025.36476682-s2.0-105025940921Q1WOS:001651993200009Q20000-0002-1033-1597