A new data-driven probabilistic fatigue life prediction framework informed by experiments and multiscale simulation

Abstract

Traditional probabilistic fatigue life test requires long time, a large number of samples and high cost due to dispersion, randomness and complexity. A new data-driven probabilistic fatigue life prediction framework informed by experiments and multiscale simulation, was proposed and applied to a Ni-based superalloy FGH96. It exhibits the advantages of less tests and fully consideration of the effects of microstructure dispersion on fatigue crack initiation, small fatigue crack (SFC) and long fatigue crack (LFC) propagation. The fatigue crack initiation life was accurately predicted by a stored energy model of persistent slip band (PSB) based on microscale crystal plasticity (CP) theory. The usage of artificial neural network (ANN) leads to the capability of rapidly obtaining CP parameters through the macroscopic Ramberg-Osgood (RO) relationship. Bayesian neural networks (BNN) was trained to predict probabilistic fatigue crack initiation life, SFC and LFC growth rates based on results obtained from CP modellings and carefully designed experiments. With BNN as a probabilistic distribution generator, a Monte Carlo (MCs) model is developed to capture the complete fatigue failure processes, including the crack initiation, SFC and LFC propagation. The total fatigue life and the proportions of life at different stages of superalloy FGH96 were predicted by MCs, which agree well with the experimental data. This work provides a new pathway for structural safety assessment and damage tolerance design.