Fatigue Growth Prediction of Interacting Cracks Using Machine Learning Nested in Finite Element Analysis

Accurately predicting fatigue crack growth in interacting cracks remains a major challenge in fracture mechanics, as empirical laws such as Paris’ law cannot adequately capture observed behaviour. This study introduces a hybrid finite element-machine learning framework that nests an artificial neural network (ANN) directly within an extended finite element method (XFEM) environment to overcome these limitations. The ANN is first pretrained on synthetic data generated from a thickness-modified Paris’ law, ensuring consistency with established physical trends. Crucially, the ANN is subsequently retrained while nested inside the XFEM simulation using limited experimental measurements, enabled by a tailored backpropagation algorithm that bypasses the need for analytically tractable derivatives. This nested training approach allows the model to adapt directly to observed crack length data rather than relying solely on empirical growth-rate equations. Validation against experimental fatigue tests on aluminium alloy 5005 with collinear cracks demonstrates that the retrained ANN substantially improves prediction accuracy, capturing both the influence of plate thickness and the existence of a stress intensity threshold below which crack growth ceases. The results demonstrate the feasibility of replacing conventional fatigue crack growth laws with a retrainable ANN nested in commercial finite element software, providing a flexible foundation for modelling complex crack propagation problems in engineering structures.