Face-Wise Prediction of Sheet-Metal Drawability Using Graph Neural Networks

The early design phase of deep-drawn structural components involves time-consuming iterative development. Traditional drawability assessments rely on finite element simulations, which are computationally expensive and slow the design process. Alternative machine learning (ML) approaches show promise in accelerating this process but face challenges with existing methods. Existing low-dimensional ML models only provide global predictions without identifying specific geometric regions prone to failure. High-dimensional models provide local predictions but require significant amounts of training data. We propose a data-driven approach leveraging graph neural networks (GNNs) for face-wise drawability prediction of sheet metal components in their computer-aided design (CAD) representation. Our method aims to bridge the gap between the computational efficiency of ML and the spatial resolution of simulation by providing face-wise insight into potential failure regions. This study utilises a dataset of parametric U-channel geometries with variability in both geometry and topology. Ground-truth labels are generated using inverse analysis simulations. Geometric entities are represented through the use of UV parameterisations, whereby 3D surfaces are mapped into 2D space to facilitate geometric encoding. Concurrently, the topological relationships are captured using a face adjacency graph. To address data scarcity, we evaluate how different amounts of training data affect model performance and perform ablation studies to analyse the impact of different CAD representation features. Our results show that the proposed approach achieves high accuracy even with limited training data. In addition, the ablation studies provide insights into the most critical CAD features, guiding future research. These results highlight the potential of our GNN to predict face-wise drawability in the early design phase.