Toughness from Imagery: Extracting More from Failure Analysis using Deep Convolutional Neural Networks

Understanding the origins of mechanical failures is critical to the prevention of future failures. In this study, additively manufactured Charpy bars, commonly used to measure the impact toughness of materials, were produced over a wide range of process conditions. The Charpy V-Notch toughness was measured on over 200 samples alongside corresponding optical images of both sides of the fracture surface. Convolutional neural network models were trained to correlate the fractographic images with quantitative toughness values. Several different neural network architectures were compared, along with other strategies for data cleaning and downsampling. The best models predicted Charpy toughness values from imagery with a mean absolute percent error of 8.5%. The neural network results were interpreted through a Grad-CAM saliency map; toughness values were correlated with expected physical characteristics such as porosity, shear lips, and fracture surface edges. A model trained on data from a Kovar alloy was found to maintain predictivity when applied to other similar alloy systems (300-series stainless steels) without any additional training. The primary optical images used in this study were macrofractography images spanning the entire fracture surface; a follow-on study using microfractographic images was less predictive but retained some utility. This work illustrates opportunities for developing data-driven approaches to provide quantitative assessment and qualitative interpretations of fracture surfaces.