Abstract Conventional fatigue life prediction methods struggle with multiaxial and non-proportional loading conditions, yet accurate prediction is critical for component reliability in aerospace, automotive, and structural applications. This paper presents a novel AI framework with physics-informed features for fatigue life prediction of metals using three advanced models: LSBoost, XGBoost, and Deep Forest. A pivotal aspect is a domain-informed feature engineering methodology that transforms raw stress–strain data into 36 physically meaningful predictive features. These features capture key aspects of fatigue behavior, encompassing loading conditions, von Mises equivalent stress and strain, Goodman-corrected mean stress effects, detailed cycle characterization, energy-based damage measures, and indicators of multiaxial stress state. A dataset encompassing 25 material types and 589 experiments spanning both stress- and strain-controlled constant-amplitude multiaxial loading conditions was employed to train and evaluate a reliable AI framework capable of accurately characterizing fatigue behavior. Model hyperparameters are tuned using Bayesian optimization with a custom composite score balancing RMSE, MAE, R 2 , and generalization gap. XGBoost achieved the highest predictive accuracy with a testing R 2 of 0.9357 (Pearson R = 0.9681) and RMSE of 0.2351 log₁₀ cycles, followed closely by LSBoost with a testing R 2 of 0.9225 (Pearson R = 0.9613) and RMSE of 0.2580 log₁₀ cycles, while Deep Forest yielded slightly lower predictive accuracy than the gradient boosting models. Permutation-based feature importance analysis identified equivalent-stress-to-UTS ratio as the dominant predictor by a substantial margin, followed by axial-shear phase difference, axial loading amplitude, and multiaxial Goodman safety factor, providing quantitative evidence that the proximity of the applied stress state to the material failure threshold and loading amplitude are the primary determinants of multiaxial fatigue life.
Seidi et al. (Wed,) studied this question.