Estimating notch fatigue limits via a machine learning-based approach structured according to the classic Kf formulas

This paper deals with the problem of estimating notch fatigue limits via machine learning. The proposed strategy is based on those constitutive elements that were used by the pioneers like Peterson, Neuber, Heywood, and Topper to devise their well-known formulas. The machine learning algorithms being considered were trained and tested using a database containing 238 notch fatigue limits taken from the literature. The outcomes from this study confirm that machine learning is a promising approach for designing notched components against fatigue. In particular, the accuracy in the estimates can easily be increased by simply increasing size and quality of the calibration dataset. Further, since machine learning regression models are highly flexible and can handle high-dimensional datasets with many input features, they can capture complex relationships between input features and the target variable. This means that the accuracy in estimating notch fatigue limit can be increased by including in the analyses further input features like, for instance, grain size or hardness. Finally, machine learning's generalization ability is crucial for regression tasks where the goal is to predict values for new materials.