An automated Machine Learning based approach for a reproducible and efficient evaluation of industrial Charpy V-notch specimens

• Machine Learning model enables evaluation of large-scale fracture surfaces. • Fully automated evaluation from sample box to individual sample. • Analysis of samples at pixel level enables highly accurate predictions. • Electron microscopy confirms results through visual overlays. An objective and automated method for the quantification of macroscopic images of tested Charpy V-notch specimens, specifically focusing on their ductility/brittleness characteristics based on a realistic, homogeneous and industrial environment is proposed. Our approach involves a multi-step preprocessing routine that incorporates color thresholding and connected component analysis to first detect the various Charpy V-notch specimen bundles according to their sample material affiliation and testing temperature. Subsequently, a U-Net was trained to further partition the preprocessed images into background, notch, and regions of ductile or brittle fracture, respectively through semantic segmentation. Thereby, a quantification of brittle and ductile fractions of each individual sample focusing on only the fracture surfaces can be conducted. The results obtained are then evaluated using Intersection over Union (IoU) metrics, a tailored domain-specific matrix and SEM images incorporating more objective annotations based on the high resolution and the higher depth of focus to assess the model performance. The findings presented in this study highlight the significant potential of machine learning and computer vision in the realm of a reproducible and objective, automated macroscopic fracture analysis on an industrial scale, providing valuable benefits for materials engineering and quality control in manufacturing processes.