An improved fatigue life prediction via defect-informed intelligent learning method

Abstract

Porosities evolving during the additive manufacturing process dominantly control the fatigue life of the fabricated component and make its prediction challenging. Generally, fracture mechanics and defect-based models are used to predict their fatigue lives, but often fail to capture the defects’ characteristics. This study introduces a new defect-based parameter derived from integrating the critical defect morphologies to improve the fatigue life prediction capability by incorporating it into machine learning algorithms. This approach effectively captures the complex relationship between the defect’s characteristics and fatigue life. The input features, such as defect size, its shape, its location, and surface roughness, were utilized to train the three popular machine learning models, namely, deep neural network, support vector machine, and random forest. The SHapley Additive exPlanations analysis indicates that defect morphology and surface roughness are among the most influential features affecting fatigue life. The results reveal that the integration of the proposed parameter into the machine learning framework enhances the prediction accuracy with an R² score of 0.84 compared to the conventional defect-based machine learning approaches with an R² score of 0.34 and the existing physics-based model with an R² score of 0.32. The proposed model predicted fatigue life that majorly falls within twice the error band. The new framework demonstrates enhanced capability in capturing the scale effect on the fatigue lives compared to the existing physics-informed machine learning models, highlighting its effectiveness and robustness. The proposed model effectively captured the complex relationship between the defect's characteristics and fatigue life. This work contributed toward unlocking a new pathway in the predictive capability of a combined machine learning and defect-based model.