A novel machine learning method for multiaxial fatigue life prediction: Improved adaptive neuro-fuzzy inference system.

• A novel machine learning method is used to predict the multiaxial fatigue life of various materials. • Non-proportionality and phase differences are used to describe the loading paths in non-data form. • Adam algorithm is used to optimize the original model to avoid getting trapped in a local optimum. • Six metallic materials are used to validate the predictive performance of the proposed model. • The generalization and extrapolation capability of different machine learning models are compared. In this study, a neuro-fuzzy-based machine learning method is developed to predict the multiaxial fatigue life of various metallic materials. First, the fuzzy inference system and neural network are combined to identify and capture the nonlinear mapping relationship between multiaxial fatigue damage parameters and fatigue life. Non-proportionality and phase differences are introduced to characterize different loading paths. Next, the Adam algorithm is employed to update the premise parameters of the original model to achieve fast and accurate convergence. Then, subtractive clustering is applied to extract fuzzy rules between input variables and output for more efficient prediction. Moreover, the hyperparameters in the proposed model are automatically optimized by the adaptive opposition slime mould algorithm to obtain the optimal model. The predictive performance of the proposed model is verified by fatigue experimental data for six materials in published literature, which indicates that the proposed model can effectively predict the fatigue life of various materials under different loading paths. Meanwhile, compared with six classical machine learning models, it is found that the proposed model has better predictive performance and extrapolation capability. [ABSTRACT FROM AUTHOR]