Application of machine learning for predicting hydrogen-assisted fatigue crack growth

Hydrogen is one of the key drivers for the transition to a zero-carbon future, however, it accelerates the degradation of pipelines by a process known as hydrogen embrittlement. Predictive models can offer significant advantages in terms of cost, time, and flexibility compared to experiments. Artificial neural network and random forest models are developed to predict hydrogen-assisted fatigue in three API 5L steels: X52, X70, and X100. The models utilize critical variables such as hydrogen pressure and stress intensity factor, obtained from experiments, as inputs to model their relationship with hydrogen-assisted fatigue crack growth. The efficacy of the models is validated, tested, and compared with experiments as well as phenomenological models. The outcome of the study reveals that machine learning, particularly artificial neural networks, can learn and subsequently predict the hydrogen-assisted fatigue in pipeline steels with good accuracy, including data (e.g., hydrogen pressures) that lies outside of the training dataset. • Machine learning models developed to predict H-assisted fatigue in pipeline steels. • Hydrogen pressure and stress intensity factor are used as model inputs. • Models are verified, and their interpolation tested under unseen conditions. • Neural networks showed superior capability in learning and generalizability.