Combining Experiments and Models via a Bayesian Network Approach to Predict Short Fatigue Crack Growth
Résumé
Identifying the short crack driving force of polycrystalline engineering alloys is critical to correlate the inherent microstructure variability and the uncertainty in the short crack growth behavior observed during stage I fatigue crack growth. Due to recent experimental advancements, data of a short crack propagating at the relevant length scale is available via phase and diffraction contrast tomography. To compute the micromechanical fields not available from the experiment, crystal plasticity simulations are performed. Results of the experiment and simulations are combined in a single dataset and sampled utilizing non-local mining technique. Sampled data is analyzed using a machine learning Bayesian Network framework to identify statically relevant correlations between state variables, microstructure features, location of the crack front, and experimentally observed growth rate, in order to postulate a data-driven, non-parametric short crack driving force.