Construction of Fatigue Criteria Through Positive‐Unlabeled Learning

Abstract The reliability of vehicles is a significant issue for automotive manufacturers, with mechanical fatigue representing a pivotal aspect of the design process. To accelerate the development of novel mechanical components, automotive manufacturers are increasingly employing numerical simulations with the objective of markedly reducing the number of prototype validation tests. This requires the utilization of efficient fatigue criteria capable of accurately identifying critical zones in numerical models. However, the current fatigue criteria frequently fail to demonstrate a satisfactory correlation with the results obtained from fatigue test rigs. In response, this paper proposes a probabilistic Dang Van criterion that accounts for variability in fatigue results within a multiaxial context. Additionally, a fatigue database comprising numerical results and test reports on automotive chassis components is introduced. A novel approach utilizing positive‐unlabeled learning is developed to enhance the predictive accuracy of the fatigue criterion. Its effectiveness is demonstrated through application to the fatigue database.