Hydrogen-assisted cracking: A deep learning approach for fractographic analysis

Hydrogen handling equipment suffers from interaction with their operating environment, which degrades the mechanical properties and compromises component integrity. Hydrogen-assisted cracking is responsible for several industrial failures with potentially severe consequences. A thorough failure analysis can determine the failure mechanism, locate its origin, and identify possible root causes to avoid similar events in the future. Post-mortem fractographic analysis can classify the fracture mode and determine whether the hydrogen-metal interaction contributed to the component's breakdown. Experts in fracture classification identify characteristic marks and textural features by visual inspection to determine the failure mechanism. Although widely adopted, this process is time-consuming and influenced by subjective judgment and individual expertise. This study aims to automate fractographic analysis through advanced computer vision techniques. Different materials were tested in hydrogen atmospheres and inert environments, and their fracture surfaces were analyzed by scanning electron microscopy to create an extensive image dataset. A pre-trained Convolutional Neural Network was fine-tuned to accurately classify brittle and ductile fractures. In addition, Grad-CAM interpretability method was adopted to identify the image regions most influential in the model's prediction and compare the saliency maps with expert annotations. This approach offered a reliable data-driven alternative to conventional fractographic analysis.