In the scope of the BMWK-funded "HL-Blech" project, various high-energy welding processes were examined to develop low-alloy steels for monopile construction. Fractography, an established method, evaluates the mechanical properties, including the brittle-ductile transition temperature. Experts analyze macroscopic images of fracture surfaces of Charpy V-notch specimens to identify ductile and brittle fractions. However, this assessment often relies on subjective judgments, posing reproducibility challenges crucial for material approval under the EN 10025-1 standard.

This study concentrates on machine learning-based semantic segmentation of Charpy V-notch specimen fracture surfaces in an industrial context. Building on recent advancements in deep learning for macroscale fracture surface segmentation, the study focuses on direct sample extraction and characterization from digital images. Using color thresholding and connected component analysis, sample bundles are identified and extracted. DenseNets and ResNets characterize the fracture surfaces in terms of ductility and brittleness. Evaluation metrics like Intersection over Union (IoU) and confusion matrices assess the results and detect potential biases within the models.

Preliminary investigation shows a satisfactory alignment between brittle, ductile, notch, and background classes in both ground truth and predictions. Despite challenges from defocused image areas, IoU (Intersection over Union) values exceeding 75% for all classes demonstrate the significant potential of this approach. In summary, our methodology provides a streamlined industrial-scale approach to macroscopic sample evaluation.