Fretting fatigue is a complex mechanism of material damage and a common cause of failure in mechanical components, yet its life prediction remains challenging for traditional empirical models. These methods typically show significant scatter and are often limited to individual material alloys. This study develops, implements, and validates data-driven neural network models capable of accurately predicting fretting fatigue life across a wide spectrum of high-strength engineering alloys and varied contact geometries. The models are trained on an experimental dataset that includes eight distinct alloys (including Aluminum, Titanium, and Inconel) and both spherical and cylindrical contact geometries. A key methodological innovation is the systematic integration of k-fold cross-validation to mitigate the risk of overfitting, a common issue with limited fretting fatigue datasets. This validation protocol maximizes data utility. We compare two complementary NN architectures: NN1, which uses only readily measurable global experimental parameters for rapid, simple deployment; and NN2, which incorporates physics-based subsurface stress–strain descriptors as engineered inputs to evaluate whether they improve prediction vs global parameters. The resulting suite of cross-validated, multi-material neural network predictors provides a tool for the rapid estimation of fatigue life under fretting conditions, supporting informed material selection and guiding experimental design in high-performance applications.
Teruel et al. (Wed,) studied this question.