Physics‐Informed Multisource Domain Transfer Learning for Fatigue Life Prediction of IN718 Alloy

ABSTRACT A physics‐informed multisource domain transfer learning (PITL) prediction model was proposed to predict fatigue life of IN718 superalloy under complex conditions, which faced sparse data and complex mechanism challenges. The model effectively integrates physical prior knowledge and transfer learning algorithms widely used in current fatigue research and engineering fields, such as equivalent effect variable model, Manson–Coffin model, and other physical information, as well as low‐cycle fatigue hysteresis‐loop characteristic information, based on its architecture. It effectively solves the problem of learning and utilizing multisource data in fatigue testing and achieves accurate prediction of fatigue life under different strain‐ratio loading conditions based on known fatigue loading conditions. And then, the effects of different model combinations, multisource transfer patterns, and multisource data size on the accuracy of target domain life prediction are systematically explored under four strain ratios loading condition. The results show that the PITL model incorporating physical information exhibits higher prediction accuracy and robustness than the basic transfer learning model. Its predictive performance depends on the model combinations, the transfer patterns, and the amount of data in the source domain.