Scaling finite element method (FEM) based corrosion simulations to whole body‐in‐white structures lead to extremely high computational costs. As corrosion only appears in corrosive critical areas, the FEM is restricted to these. The objective is a semantic segmentation of corrosive critical designs, which are part edges and flanges in body‐in‐white structures. Different deterministic as well as Machine Learning and Deep Learning approaches are proposed and compared with respect to their ability to segment critical designs based on the geometry and the spatial relations of the parts only. The deterministic edge detection provides a fast and highly accurate way to segment part edges, whereas the described deterministic flange detection is not suitable for capturing the full diversity of flange structures. As the feature‐based Machine Learning approach evaluates more properties in a more flexible way, the performance of the flange detection is significantly increased and even the edge segmentation obtains slightly better results. The evaluation of graph structures with a Geometric Deep Learning method fails as the train set is too small to sufficiently represent the complex and various structures in the test set.

The edge corrosion simulation of Kapfer et al. is used to extract time-dependent delamination widths of metal sheet edges with different spatial orientations, encoded by edge classes and angular displacements with respect to the ground.The corrosion behavior of the edges of a body-in-white part is examined with respect to the spatial orientation and compared to the simulated delamination widths, demonstrating that the edge corrosion simulation is also able to model more complex structures. Since scaling the finite element method-based corrosion simulation to the whole part requires a lot of memory and computational time, different neural network types are evaluated to predict the delamination widths. Together with the geometric properties of the part edges, a fast and accurate prediction of edge delamination of body-in-white parts is possible, requiring few time-consuming corrosion simulations only in the training phase of the networks.