Nix Award and Lecture Symposium V

About the Presentation
Crack initiation governs high cycle fatigue life and is sensitive to microstructural details. While microstructure-sensitive models are available, their validation is difficult. We have therefore developed a combined experimental and data post-processing workflow to establish multimodal observation of fatigue crack initiation and propagation efficiently. It involves fatigue testing of mesoscale specimens, data fusion through multimodal registration, and image-based data-driven damage localization. We then propose a validation framework where a fatigue test is mimicked in a sub-modeling simulation by embedding the measured microstructure into the specimen geometry and adopting an approximation of the experimental boundary conditions. This simulation-based approach is compared to training graph convolutional networks on the single grain level. Such graph convolutional networks yielded the best performance with a balanced accuracy of 0.72 and a F1-score of 0.34, outperforming the phenomenological crystal plasticity simulations and conventional machine learning models by large margins.