A Physics-Guided Neural Network for Probabilistic Fatigue Life Prediction Under Multiple Overload Effects

A Physics-guided Neural Network (PgNN) is proposed to provide a robust probability distribution of fatigue life under arbitrary multiple overloads, which integrates the physical mechanism model (PMM) and neural network (NN). Notably, the proposed PgNNs are trained solely using data under constant amplitude loading scenarios. Firstly, a PMM is developed to predict fatigue life based on linear elastic fracture mechanics, considering crack closure. A data preprocessing approach informing PMM is presented, transforming arbitrary overload conditions into equivalent constant amplitude loading with stress ratio R=0$$ \mathrm{R}=0 $$. Moreover, a back-propagation NN is constructed, where a loss function integrating the PMM and mean square error is designed. The PgNN framework encompasses the uncertainties associated with stress levels, material coefficients and equivalent initial flaw size. The fatigue data of aluminum alloy 7075-T6 and Al-Li alloy 2060 are used for model validation. The results affirm that the PgNN exhibits superior accuracy and robustness.