Application of the Gaussian process for fatigue life prediction under multiaxial loading

Abstract

Fatigue life prediction with failure probability estimation for materials subjected to multiaxial loading is an important task in engineering design because it reduces the financial cost by eliminating expensive experimental tests. Owing to the complex deterioration mechanisms of fatigue failure, existing fatigue life prediction models are of the empirical or semi-empirical type, and their applicability is limited to validating the loading condition and material. Improper selection of the fatigue model could result in non-conservative life prediction with high financial and catastrophic consequences. To solve this problem, an innovative approach based on the Gaussian process for fatigue life prediction under multiaxial loading is presented. The inherent features of the Gaussian process predispose its application to fatigue life prediction under multiaxial loading as an efficient and practical approach to avoid the problem in selecting adequate semi-empirical parametric fatigue models. The model was validated on two sets of experimental data obtained by fatigue testing of S355N steel and 2124 T851 aluminum alloy under uniaxial and multiaxial loadings. In the model training process, only data obtained under uniaxial and pure torsion cyclic loadings were applied. Owing to the implemented physics-based input data, which are related to the stress components acting on the critical plane of crack initiation, the model accurately predicted the fatigue lives of the two tested materials by implementing the squared exponential covariance function. In addition, the fatigue lives were also computed using four parametric fatigue strength criteria (Crossland, Matake, Carpinteri et al., and Papuga–Růžička models). The best results obtained by the parametric models exhibited a lower fatigue prediction performance than the results using the Gaussian process-based model.