CC BY 4.0Bauer, Niklas MichaelNiklas MichaelBauerBaumgartner, JörgJörgBaumgartner2025-07-022025-07-022025-06-03https://doi.org/10.24406/publica-4839https://publica.fraunhofer.de/handle/publica/48908810.1007/s40194-025-02080-910.24406/publica-4839Welded joints made of ductile materials exhibit a significant reduction in fatigue life under non-proportional multiaxial loading compared to proportional loading. Different methods to calculate multiaxial fatigue life are evaluated on a comprehensive database, which is created based on various fatigue tests showing ductile material behavior under uniaxial, proportional, and out-of-phase loading. This paper finds that well-known stress-based methods widely used in the literature to evaluate multiaxial fatigue of welded joints cannot accurately calculate fatigue life under both proportional and non-proportional loading for welded ductile materials. Two new methods providing reliable and accurate fatigue life predictions with little scatter under both constant and variable amplitude loading are developed, analyzed in detail, and compared to existing approaches. The new methods include an interaction equation derived from the criteria proposed in codes and standards, as well as a machine learning approach based on artificial neural networks, both developed and optimized or trained based on the created database. Using interpretable machine learning, the neural networks are found to have learned similar correlations between multiaxial loading and fatigue life as those represented by the new interaction equation.enFatigue life predictionNon-proportional loadingWelded jointsMachine learningjournal article