Applying Machine Learning and Computer Vision to the Study of Fatigue Fracture Surfaces

  Images of fracture surfaces are rich in information and have long been examined to determine  fracture characteristics such as initiation site and crack growth mode. Computer vision and  machine learning techniques, including convolutional neural networks (CNNs), are novel  methods that can be applied to microstructural images to make predictions and connections  between the microstructure and material responses. For fatigue failure, the goal is to connect  the visual appearance of the fatigue fracture surface to fatigue characteristics such as loading  values, crack length, and crack growth rate. This allows quantitative information to be  obtained from fracture surfaces where specific loading conditions are unknown. Scanning  electron microscopy (SEM) has been the primary image modality for studying fracture  surfaces, but surface height information acquired from Scanning White Light Interferometry  (SWLI) can provide additional useful information. Backscattered electron (BSE) images  provide compositional information that can be used to differentiate phases in a material  and help determine how they influence the fracture surface.  

This project collects multimodal data from Ti-6Al-4V fracture surfaces in the form of  Secondary Electron and BSE images and SWLI height data and combines them to train  a CNN and identify high stress points, crack initiation sites, and predict values such as  stress intensity factor, crack length, and crack growth rate. Through data fusion, these  image modalities can be combined to improve the recognition of subtle features on the  fatigue fracture surface. Using standard computer vision techniques to register images and  transfer learning to construct CNNs, the stated quantitative values can be predicted. The  images used to develop this model, the image registration technique, the creation of CNNs,  identified fatigue properties, and fracture characteristics will be presented.