Mooring systems for floating offshore wind turbines and FPSOs experience stochastic, direction-dependent cyclic loading, yet conventional coupled hydrodynamic–finite-element fatigue workflows are far too slow for continuous digital-twin updating. We present FiGNNet, a fracture-informed graph neural network that represents a six-line spread-mooring system as a 37-node graph and predicts nodewise mixed-mode stress-intensity range directly from metocean descriptors and local structural features. Two graph-convolutional message-passing layers with a linear skip path are trained under a composite loss enforcing consistency with the Paris–Erdogan crack-growth relation, and a Wiener-process degradation layer propagates uncertainty through Monte Carlo crack-growth simulation. The method is validated on physical experimental data from the FOCAL Campaign 2 wave-basin tests. Across five instrumented records under 5-fold cross-validation, FiGNNet recovered equivalent stress-intensity range (R² = 0.52) and damage rate (R² = 0.53) from measured tension histories, with leave-one-case-out testing confirming generalization across wave-driven sea states. On a complementary 120-scenario numerical benchmark (4320 node-level samples), the model reached R² = 0.688 and RMSE = 1.86 MPa√m for stress-intensity range, and its Monte Carlo critical-node life estimates agreed with reduced-order-model references at R² = 0.81.Critically, a full-graph forward pass executes in approximately 94 μs ,six to seven orders of magnitude faster than the coupled hydro-FE evaluation it replaces real-time, uncertainty-aware fatigue tracking of an entire mooring array computationally feasible for the first time. Current limitations are the small experimental sample and degraded performance on zero-wave decay cases, which are being addressed through an expanded basin-test campaign. FiGNNet offers a reproducible, physics-constrained route from measured mooring response to actionable remaining-life estimates for offshore digital twins.