A new fatigue life prediction method for welded joints based on machine learning incorporating defect information and physics of failure

This study is dedicated to establishing a new fatigue life prediction model for welded joints that comprehensively considers the effect of welding defects and the physical mechanism of fatigue failure. Firstly, an improved fatigue model for welded joints based on damage tolerance concept is proposed, in which fatigue-critical factors such as the geometric characteristics and location of welding defects are fully considered. Secondly, the proposed physical model is incorporated into a popular machine learning approach of support vector regression (SVR) that is suitable for training fatigue test data, such that a physics-guided artificial intelligence (AI) method is established. The developed method was then applied to predict the fatigue life of welded joints of GH4169 nickel-based superalloy under different fatigue testing conditions at high temperatures. The experimental results show that the life prediction accuracy of the proposed physics-integrated SVR method is significantly higher than that of purely physics-based or purely data-driven methods. Moreover, the proposed method also exhibits better prediction capacity and applicability compared with other physics-AI hybrid methods for fatigue assessment of defect-containing welded joints. This is supported by the fatigue mechanism confirmed by fractographic analysis, demonstrating the promising potential of this physics-integrated AI approach in fatigue life prediction.