A general physics-informed neural network framework for fatigue life prediction of metallic materials

The prediction of fatigue life in metallic materials using machine learning faces challenges related to interpretability and generalization. Motived by this, a physics-informed neural network framework for fatigue life prediction is proposed in this study. Physical constraints are incorporated into the framework through partial differential inequalities derived from experimental observations. Fatigue experiments with LZ50 steel and SLM 316L are newly designed, along with experimental data for nine different metallic materials taken from literature, to validate the predictive accuracy and generalization of the model. K-fold cross-validation and Bayesian methods were employed to optimize the model architecture and physical constraint weights. The results are compared with those obtained from a traditional artificial neural network . It is demonstrated that the physics-informed neural network achieves higher predictive accuracy, better consistency with physical principles, and improved generalization in most cases. More reliable predictions are observed in high-life regions, where artificial neural networks tend to exhibit greater variability. These findings highlight the versatility and robustness of the proposed framework, which is suggested as a promising general approach for fatigue life prediction through the integration of physical knowledge into machine learning-based modeling.