Machine learning and physics based empirical equations for fatigue life prediction of AM Ti-6Al-4 V

Predicting fatigue life in additively manufactured (AM) Ti-6Al-4 V is constrained by limited datasets and the complex interplay of stress, defect attributes, and build orientation. Existing attempts to address data scarcity through synthetic augmentation have rarely validated the fidelity of generated data. This study systematically benchmarks multiple augmentation strategies under rigorous fidelity constraints to ensure preservation of the stress-defect-life relationships before model training. The Residual Bootstrapped Conditional Regression Sampler (RB-CRS) is identified as the most reliable augmentation strategy for model development. Among the four regression models tested, a compact artificial neural network (ANN) yielded the highest predictive accuracy with a test R² of approximately 0.98. SHapley Additive exPlanations (SHAP) analysis revealed that stress amplitude accounted for 74% of the predicted variation, build orientation 16%, defect size 7%, distance to surface 2%, and defect depth 1%. The ANN is deployed as an interactive tool for rapid fatigue life estimation from stress, orientation, and defect inputs, enabling rapid screening within the model’s domain. In parallel, orientation-specific, physics-guided Basquin-type equations were derived, which accurately tracked experiments up to 10⁸ cycles (R² 0.94–0.95). This integrated approach enables reliable fatigue-life prediction and supports qualification of AM Ti-6Al-4 V.