A Digitalized Machine Learning-Assisted Workflow to Study Microstructural Fatigue Damage Evolution and Crack Initiation
A characteristic crack formation process ranging from accumulation of plasticity in individual grains over the formation of extrusions and micro crack initiation to ultimately short crack formation and failure determines the fatigue lifetime of a material. The early stages in particular are determining lifetime when cyclic loading with low amplitudes is applied.
The presented work utilizes a fatigue damage evolution testing methodology, which is able to sensitively detect fatigue states up to short crack growth. This pronounced sensitivity is achieved by sample miniaturization and optimized control mechanisms. Simultaneously, images are aquired in-situ with a bidirectional stoboscope illumination allowing localization of damage development. Extending upon this work, a framework was built that spatially aligns these image time series with micro texture and high resolution scanning electron microscopy (SEM) information before and after fatigue by multi modal data registration. Furthermore, a U-Net deep learning model operates on the SEM images after fatigue to automatically segment early degradation instance like extrusions and micro cracks.
This multi modal data is post processed to contain microstructure and loading dependent features and fatigue time series derived labels. A random forest approach is followed to statistically model damage initiation sites in the microstructure and therefore provides a link between microstructure and damage initiation property. The capability of the data-driven ML model to predict extrusion and crack initiation is assessed and the workflow is presented.