Assessment Of Fatigue Crack Propagation Life Based On Physics-Informed Neural Networks Under Ensemble Learning In Hydrogen Environments

Abstract

Hydrogen storage and transport systems use pipeline steel, where hydrogen embrittlement accelerates fatigue crack propagation and reduces service life. Accurate life evaluation is necessary for safe operation. Conventional approaches experimental fatigue testing, finite-element simulation, and empirical laws such as the Paris model, are respectively costly, dependent on often-unavailable hydrogen constitutive laws, and deterministic with no uncertainty quantification. This paper offers a PINN and ensemble learning framework. The method integrates physical and data-driven modeling using Bayesian Model Averaging with a probabilistic neural network. Validation on API-X70 and X100 steels BMA-PINN achieves a log MSE of 0.0538 and R 2 of 0.9934 for X100, outperforming other models. PINN captures fatigue life stochasticity with 95th-percentile intervals.