AIDH2
The project aims to develop a deep learning based surrogate model for phase field damage predictions due to hydrogen embrittlement in steel. While phase field models elegantly capture complex phenomena like nucleation, propagation, and merging of cracks, they suffer from issues related to parameter selection and computational time. Thus, a computationally efficient machine learning approach will be explored to bypass expensive computations and provide instantaneous results towards revolutionizing hydrogen materials design and application. This will enable real-time assessment of material performance, facilitate the development of steel parts with superior resistance to hydrogen embrittlement, and drive advancements in safety-critical sectors such as aerospace, automotive, and energy, ultimately leading to more resilient and reliable infrastructure.
The physics-informed deep learning model considers the impact of steel microstructural features such as phase distribution, grain boundary structure, material parameters, crack propagation, hydrogen concentration, and stress profiles. By training with a physics-informed loss function, it is ensured that the deep learning model respects the known physical laws, enhancing accuracy and reliability of the predictions. The developed physics informed deep learning model’s validity will be studied through the k-fold cross validation method to assure that the model can provide accurate results for application cases that the deep learning model has not been trained on. This is followed by the evaluation of the feature significance and feature interaction through the shapely additive explanations to perform a parameter study and sensitivity analysis, which will offer insight into the selection of parameters and study the interaction between the stresses, hydrogen concentration and damage in the domain.
To create such a model, two different expertise are required. The knowledge of modelling hydrogen transport in steels, through the different microstructural features and enhancing the phase field model to accommodate damage due to hydrogen embrittlement is offered by the Helmholtz-Zentrum Hereon’s team. Moreover, the Hereon team will contribute its expertise in the application of uncertainty quantification (UQ) for the validation of the developed surrogate model. The competence of University of Alberta’s team includes developing machine learning models to tackle challenges in the target research area, which encompasses the optimization of the physics informed loss function, training of the network and the validation of the developed machine learning model. Only by bringing the competence of these two excellent research institutions, developing such an ambitious model is possible.
Contact
Department of Interface Modelling