Prediction of Fatigue Life of Al2O3 Using Machine Learning Algorithms

Abstract

Fatigue failure is a major cause of structural degradation in engineering materials subjected to cyclic loading, particularly in aerospace, automotive, biomedical, and manufacturing applications. Although aluminum oxide (Al₂O₃) ceramics exhibit excellent hardness, thermal stability, wear resistance, and corrosion resistance, accurately predicting their fatigue life remains challenging because of the complex nonlinear interactions among loading conditions, material microstructure, and crack evolution. This study proposes a comprehensive machine learning (ML)-based framework for predicting the fatigue life of Al₂O₃ using conventional, ensemble, and deep learning algorithms. The novelty of the study lies in the comparative evaluation of multiple ML models using fatigue-related parameters to identify the most reliable predictive approach while also revealing the material variables that govern fatigue behavior. Model performance was assessed using Mean Squared Error (MSE), Root Mean Square Error (RMSE), Mean Absolute Error (MAE), and the coefficient of determination (R²). The results show that deep learning and ensemble models significantly outperform conventional regression techniques. The Long Short-Term Memory (LSTM) model achieved the highest predictive accuracy with an RMSE of 0.108, MAE of 0.088, and R² = 0.98, followed by XGBoost (RMSE = 0.116, MAE = 0.093, R² = 0.97). Feature importance analysis further revealed that stress amplitude and temperature were the dominant predictors of fatigue life, consistent with established fatigue crack initiation and propagation mechanisms in ceramic materials. These findings demonstrate that advanced ML algorithms not only provide highly accurate fatigue life predictions but also identify the key physical variables governing fatigue degradation in Al₂O₃.