Data-driven machine learning approach for predicting dwell fatigue life in two classes of Titanium alloys

Abstract

Predicting dwell fatigue life accurately in Titanium (Ti) alloys is vital for designing reliable components such as aero-engine discs and blades. This study proposes a data-driven machine learning (ML) approach for precisely predicting dwell fatigue life in two classes of Ti alloys: near α and α + β alloys. To build a robust model, the data-driven approach tackles this multi-dimensional problem by training a comprehensive dataset of various input features, including alloy composition, monotonic properties, and fatigue testing parameters. Advanced ensemble-based algorithms, Gradient Boosting (GBOOST) and Extreme Gradient Boosting (XGBOOST) are employed. The split-insensitive best-performing XGBOOST model achieved a high accuracy (R2) score of 89.9 %, indicating impressive prediction capability. The model’s robustness is further assessed using an unseen validation dataset to draw qualitative and quantitative conclusions. ML model’s validation further involved comparing its predictions with actual experimental trends across relationships, such as dwell life versus dwell time and dwell-debit versus dwell time. Dwell-debit is also predicted as a function of varying stresses normalized to 0.2 % yield strength, highlighting the susceptibility of investigated Ti alloys to dwell fatigue. Additionally, the SHapley Additive exPlanations (SHAP) analysis was conducted to gain deeper insights into various input features affecting dwell fatigue life.