Predicting Metal Fatigue Life Under Multiaxial Loading Using Machine Learning: A Comparative Performance Analysis

Abstract

The aim of this research is to increase the accuracy of multiaxial fatigue life prediction using several machine learning models and examining their results. The dataset used in this study consists of 1,167 data points of 40 distinct materials which include some important mechanical parameters and load cases. Support Vector Machine (SVM), Random Forest (RF), Decision Tree Regressor (DTR), AdaBoost and Stacking Regressor were among the studied machine learning models. To begin with, the models were subjected to training and testing on an 80:20 data proportion. Hyperparameter tuning using GridSearchCV was performed for SVM and DTR as their R² scores were relatively low. Moreover, the paired t-tests were applied in order to evaluate the difference in the models statistically, while the prediction intervals were calculated in order to illustrate the level of assurance on the predictions. Among all models, the Stacking Regressor achieved the highest R² score of 0.84, outperforming RF, DTR, and AdaBoost. Its superior performance demonstrates its potential for real-world applications in predicting fatigue life, such as enhancing reliability in mechanical systems and reducing costs associated with experimental testing. This study underscores the value of ensemble techniques like stacking in addressing complex engineering problems and improving prediction accuracy.