Fatigue Life Prediction in 7075 T651 Aluminum Alloy: Effects of Applied Stress

Abstract

This study examines the relationship between applied stress (σa [MPa]) and fatigue life (Nf) in 7075-T651 aluminum
alloy using both physics-based and machine learning approaches. Fatigue tests were conducted under bending and
torsional loading at multiple time stages, and the fatigue life was modeled on a log10(Nf) scale. A Basquin-type
regression (log10(Nf) ~ stress amplitude) captured the expected monotonic stress–life trend, explaining approximately
one-third of the variance (R² ≈ 0.33) with an RMSE of 0.71 log cycles (≈3.2 × 10⁶ cycles) and an MAE of 0.58 log cycles
(≈1.1 × 10⁶ cycles). An optimized XGBoost model trained with applied stress and time stage reduced the absolute error
(MAE ≈ 7.6 × 10⁵ cycles) but accounted for less variance overall (R² ≈ 0.17; RMSE ≈ 0.84 log cycles, ≈2.4 × 10⁶ cycles).
Feature importance analysis confirmed that applied stress was the dominant predictor, with time stage contributing
only marginally. These findings show that while XGBoost achieved modest improvements in raw error, the Basquin
regression remained more interpretable and physically consistent, underscoring the importance of benchmarking ML
models against established fatigue mechanics.