Predicting fatigue crack growth metrics from fractographs: Towards fractography by computer vision.

This work utilized computer vision and machine learning techniques to predict both qualitative characteristics and quantitative values, from SEM images of Ti–6Al–4V fracture surfaces from compact tension specimen fatigue crack growth tests. This work found that Convolutional Neural Networks (CNNs) focused on different features in images based on the length scale of the image. This study determined a lower limit field of view related to the number of grains imaged, and confirmed that transfer learning of a pre-trained CNN can distinguish between two forging direction and two different load ratios, and predict crack length, a , and repurposed for Δ K , and d a d N. • A pre-trained CNN classified forging direction based on fatigue fracture surfaces. • The same CNN modified for regression predicted crack length from only fracture images. • Longer length scale features capturing multiple grains are better for models. • Models can predict crack growth rate stress intensity factor range, and load ratio. [ABSTRACT FROM AUTHOR]