Machine learning prediction of stress intensity factors of corner cracks in adjacent holes

Purpose This study addresses the high computational cost and complexity of conventional numerical methods for predicting the geometric correction factor (β-factor) associated with the stress intensity factor. It introduces a machine learning (ML) framework that enables accurate, efficient, and interpretable predictions. Model interpretability is explored using SHAP to quantify feature importance and generalization is assessed across various loading conditions, with a focus on bending. Design/methodology/approach Five ML models: support vector regression, k-nearest neighbors, extreme gradient boosting (XGBoost), artificial neural network and random forest , were evaluated for predicting the β-factor of oblique part-elliptical corner cracks in riveted lap joints. Hyperparameters were optimized using particle swarm optimization, and model performance was assessed using standard regression metrics via cross-validation. The best-performing model was further analyzed using Shapley Additive Explanations (SHAP)to interpret feature contributions. Its generalization was tested under pin-loading and bending scenarios. Findings XGBoost outperformed other models, achieving the highest R2 value (0.9815) and the lowest error metrics (Mean Absolute Error, Mean Square Error, Root Mean Square Error), demonstrating excellent predictive accuracy and reliability in estimating the β-factor. SHAP analysis identified the crack depth-to-length ratio (a/c) as the most influential feature, while the hole radius-to-thickness ratio (r/t) had minimal impact. The model generalized well under bending but showed reduced consistency under pin-loading. Originality/value Unlike traditional numerical approaches that rely on geometric simplifications and are computationally intensive, the proposed ML framework offers fast, accurate and interpretable predictions that generalize effectively to varied loading conditions, particularly bending.