Graph Neural Network-based Generation of Stamped Bipolar Plate Geometries from Tooling Surfaces

The Finite Element Method (FEM) is the predominant simulation tool for planning and analyzing deep-drawing and stamping processes, providing accurate predictions of final part thickness, shape, and material failure. However, these detailed predictions are computationally expensive, especially for complex forming processes involving large geometries. The one-step approach offers a computationally efficient alternative but requires prior knowledge of the final part geometry, which limits its applicability. We introduce a Graph Neural Network (GNN)–based surrogate model that predicts the final shape of sheet metal components for a given tooling geometry by displacing nodes from a flat blank. The predicted geometry can then serve as input for a one-step solver, enabling fast manufacturability assessments. Our trained model demonstrates the ability to generate reasonable part geometries from tool surfaces. The strains and thinning resulting from a subsequent one-step analysis also show good agreement with a reference geometry.