The WBBT, as described in Stahl-Eisen-Prüfblatt (SEP) 1390 [1], is used to demonstrate that structural steels intended for use in safety-critical applications, e.g. those erected in accordance with ZTV-ING Part 4 [2], have a sufficiently high crack arrest capacity. Three possible test results are distinguished: passed (p) = α ≥ 60 ° without fracture; not passed (n.p.) = fracture at α < 60 °; invalid = no crack recognisable in the base material after bending until α ≥ 60 °. In order to reduce material usage, emissions, and the time required for testing, this study proposes the use of machine learning (ML) to predict WBBT results. Various material-related analyses were carried out to justify the implementation of the model's input variables (features). For instance, microstructure analyses revealed a correlation between the delivery condition of the material, its microstructure, and its failure behaviour. Notably, the Widmannstätten ferrite, which had a slightly coarser grain structure, was particularly prevalent in the n.p.-class. In contrast, the p-class predominantly exhibited a finer ferritic-pearlitic microstructure. Consequently, an extensive database has been set up in collaboration with CEWUS GmbH. This covers 2,610 WBBT samples and 26 features, which are both categorical and numerical. A thorough comparison of six algorithms was carried out in the ML modeling. Notably, the Bagging Classifier based on Random Forests (BCRF) achieved a high Balanced Accuracy of 74.3 %. The most significant features for this prediction are material thickness, Z-grade, Charpy impact energy and P-content.