Hybrid clustering-enhanced interpretable machine learning for fatigue life prediction across various cyclic stages in laser powder bed fused Ti-6Al-4V alloy

Predicting the fatigue life of parts fabricated by laser powder bed fusion (LPBF) is essential for assessing their durability and reliability in mission-critical load-bearing structures. Herein, we propose a novel machine learning (ML) strategy for predicting fatigue life of LPBF parts. It combines three-stage feature screening, data augmentation, a hybrid clustering integrated regression model (H-CIRM) and SHapley Additive exPlanations (SHAP) method. The optimal number of clusters is 3, and tailored models for each cluster are gradient boosting decision trees (coefficient of determination, R² = 0.951), random forest (R² = 0.930) and extreme gradient boosting (R² = 0.877), respectively. H-CIRM outperforms separate models across various cyclic stages, achieving an overall increase of over 4 % in R² and a 22 % decrease in mean absolute error (MAE). The relative error is within ± 10 % on an additional experimental dataset. This strategic clustering effectively improves predicted accuracy for each cluster, markedly enhancing overall predictive. SHAP elucidates that stress amplitude (σ_a) and the complex interactions between density (ρ), laser volume energy density (E_V), and σ_a are the top three factors affecting fatigue life, with a relative importance greater than 20 %. This work develops a robust model for estimating fatigue life of LPBF parts and provides valuable insights into optimizing process parameters.