Analytical and machine learning-based fatigue life prediction of welded joints under multiaxial loading

• ML methods enhance fatigue prediction beyond traditional design criteria. • ANN achieved the highest fatigue life prediction accuracy. • XGBoost performs better with random data splits than programewise. • SEC provides very good results only slightly less accurate than the ML models. • SHAP analyses enable model agnostic interpretation of the data. Evaluating the fatigue life of welded joints under multiaxial loading is a key challenge in structural engineering. This study explores machine learning (ML) methods for predicting fatigue life and compares their performance against the novel super ellipse criterion, which is an analytical approach that aims to improve current design standard methods (e.g., Eurocode 3, IIW). Using a dataset of uniaxial and multiaxial fatigue tests with varying phase angles, ML models—including artificial neural networks and extreme gradient boosting (XGBoost)—are trained on features like stress amplitudes, phase differences, and material properties. Artificial neural networks provide high accuracy, while tree-based models like XGBoost offer better interpretability via model agnostic interpretation using Explainable Artificial Intelligence. Results show ML models can outperform traditional criteria, especially under non-proportional loading, but face limitations near the edges of the training data. This work highlights the potential and challenges of ML in fatigue prediction and highlights their value for enhancing the safety and reliability of welded structures.