Physics-Informed Machine Learning Framework for Fatigue Life Prediction of Additively Manufactured Alloys

The fatigue life prediction of additively manufactured (AM) alloys remains challenging owing to process-induced defects, microstructural variability, and complex loading conditions of the alloys. This study presents a domain-knowledge-informed machine learning (ML) and deep learning (DL) framework for fatigue life prediction, in which physically motivated fatigue descriptors are integrated into the feature space using experimentally obtained stress–life (S–N) data. Four physics-guided engineered descriptors, namely the normalized stress (σa/UTS), R-modified stress amplitude, UTS/YS ratio, and elastic strain energy density, were incorporated into the modelling framework to improve mechanistically grounded learning across diverse alloy systems. Five ML/DL models, namely Deep Artificial Neural Network (DANN), XGBoost, Extra Trees, Stacking Ensemble, and Random Forest, were benchmarked against the classical Basquin stress–life baseline. DANN achieved the best test-set performance (R2 = 0.7114, RMSE = 0.5205 log cycles), whereas XGBoost exhibited the highest cross-validation performance (R2 = 0.7547 ± 0.056). Ablation analysis confirmed the positive contributions of both the engineered descriptors (ΔR2 = +0.115) and runout indicator (ΔR2 = +0.107) to the predictive capability. The runout flag is appropriate for retrospective database modelling. For prospective applications, the no-runout configuration (R2 = 0.5504) substantially outperformed the Basquin baseline (R2 = 0.1244) and is recommended when runout information is unavailable. TreeSHAP analysis identified normalized stress and elongation as dominant predictors, with σa/UTS showing substantially greater importance than did the raw stress amplitude. The results demonstrate that physics-informed feature engineering substantially improves fatigue life prediction across the alloy systems and processing conditions represented in the dataset; however, further validation is required for under-represented additive manufacturing processes and alloy classes.