A physics-informed Bayesian neural network model for probabilistic prediction of fatigue crack growth rate at different temperatures

Abstract

A physics-informed Bayesian neural network (PIBNN) method is proposed for predicting the probabilistic model for fatigue crack growth rate (FCGR) and fatigue crack growth life (FCGL). First, the Bayesian neural network (BNN) is employed for the probabilistic prediction of FCGR. Subsequently, the physical FCGR model that accounts for temperature effects is integrated into the BNN model as a loss function to enhance the model’s generalization capability. The mean and dispersion errors of the predictions at different temperatures are less than 10% compared to the test data. The probabilistic prediction of FCGL is ultimately realized through the integration of PIBNN with a Markov chain, in which the state transition probability matrix characterizes the transition probabilities between dispersion states.