Physics-constrained Gaussian process for life prediction under in-phase multiaxial cyclic loading with superposed static components

• Published in: International journal of fatigue Vol. 175; p. 107776
• Main Authors: Karolczuk, Aleksander, Liu, Yongming, Kluger, Krzysztof, Derda, Szymon, Skibicki, Dariusz, Pejkowski, Łukasz
• Format: Journal Article
• Language: English
• Published: Elsevier Ltd 01.10.2023
• Subjects: Fatigue life prediction; Machine learning; Multiaxial loading; Physics-constrained Gaussian process; Multiaxial loading; Physics-constrained Gaussian process; Fatigue life prediction; Machine learning
• ISSN: 0142-1123; 1879-3452
• DOI: 10.1016/j.ijfatigue.2023.107776

•Innovative framework to construct GP model for fatigue life prediction is developed.•Fatigue-based physics constraints were incorporated into the GP model.•GP trained on one-dimensional cyclic loading cases outperformed parametric models.•Fatigue-based physics constraints imposed on GP reduce the overfitting risk. Under multiaxial fatigue loading, the superposed static components are additional factors for life prediction models to be considered. The increased dimension in fatigue data imposes difficulties in pattern recognition using existing functional form models. A framework to build a Gaussian process (GP) model for lifetime prediction under multiaxial loading was developed to solve this problem. Physically consistent constraints were imposed by applying a novel technique on the GP model to control its behavior and to decrease an overfitting risk. The model consistency with the rotationally invariant principle of damage was provided by the application of the critical plane concept. The framework was demonstrated to have excellent prediction capability on S355 steel and 7075-T651 aluminum alloy. Five well-known fatigue models of functional forms were also implemented for comparison. Detailed parametric studies were presented for the training sample effect, GP kernel effect, and model predictability.