Hybrid-driven high-cycle fatigue life prediction: Physically constrained neural network optimization with small sample learning

• Published in: International journal of fatigue Vol. 201; p. 109128
• Main Authors: Jing, Guoxi, Tao, Shuai, Ma, Teng, Fu, Yafei, Zhou, Zhige, Zhang, Liqiang
• Format: Journal Article
• Language: English
• Published: Elsevier Ltd 01.12.2025
• Subjects: Fatigue life prediction; High cycle fatigue; Physical neural network; S-N curve; S-N curve; Fatigue life prediction; Physical neural network; High cycle fatigue
• ISSN: 0142-1123
• DOI: 10.1016/j.ijfatigue.2025.109128

•Constructed a physically constrained neural network based on fatigue stress life distribution law.•A modified double-transition S-N curve model based on the Basquin equation is proposed.•Comparison of fatigue life prediction of four models in six datasets.•The effect of sensitivity of physical information neural networks is considered. The progressive application of machine learning in the field of fatigue life research has led to the development of data-driven artificial neural network (ANN) based algorithms that fit fatigue test data. However, these algorithms have produced results that are inconsistent with the commonly accepted physics of fatigue properties and are therefore deemed to be overfitted. This paper puts forth a methodology that combines the physical information neural network (PINN) to train multiple fatigue life test data sets. This methodology introduces constraints on the loss of physical information and evaluates the superiority of the fitting. The objective is to ensure that the fitted full-life S-N curves are highly accurate while satisfying the physical consistency of the materials. The findings demonstrate that the mean life expectancy predicted by the PINN model, based on the S-N curve, is unambiguous and accurate, with a higher degree of precision than that observed in the three-parameter and the double-transition S-N curve model. Furthermore, in comparison to the ANN model, the PINN model effectively circumvents the overfitting issue. The incorporation of penalty factors to alter the weight of physical information constraints significantly enhances the physical consistency of the PINN model in predicting fatigue life.