Non-planar crack propagation prediction under mixed-mode fatigue in steel jacket structures using machine learning

Fatigue at welded tubular joints often limits the service life of offshore steel jacket structures. However, detailed fracture mechanics analyses of mixed-mode fatigue crack propagation remain computationally demanding for routine design and maintenance. This study develops a multi-fidelity finite element (FE) model of a steel jacket and validates crack propagation modeling in Abaqus and Franc3D against a full-scale fatigue test on a K-joint. The corresponding fatigue life was within 8% of the experimental results, establishing a validated numerical environment for computing mixed-mode stress intensity factors (SIFs) K I and K II . In this environment, mixed-mode SIFs along non-planar crack fronts were computed for a range of crack configurations and loading conditions and used to train eight machine learning (ML) surrogates. Among these, a deep neural network (DNN) achieved the highest accuracy, with a mean absolute error of 3.2 MPa·mm 1/2 . After identifying the DNN as the best-performing model, it was applied in an incremental crack growth procedure to two cases, where it predicted crack propagation and fatigue life with less than 7% deviation from simulation results. This combined multi-fidelity and ML-based approach reduces computational time from 4 h to 3 s per analysis and supports efficient and scalable fatigue assessment of offshore structures. • An ML-based framework predicts mixed-mode fatigue crack growth in steel jackets. • Full-scale K-joint tests validate the FE modelling with fatigue life error within 8%. • DNN surrogate outperforms tree ensembles, SVR, and kNN in SIFs prediction accuracy. • DNN predicts mixed-mode SIFs with MAE of 3.2 MPa mm 1 ᐟ 2 and fatigue life within 6%. • DNN reduces fatigue evaluation time from 4 h to a few seconds per case.