Enhancing FEA sheet metal forming simulations through 3D geometric data-driven surrogate modeling

Abstract Finite Element Analysis (FEA) simulation results of sheet metal forming processes are highly sensitive to changes in component geometry, as any alteration requires a complete re-computation of the forming process. Surrogate models, trained on FEA simulation data, offer a promising alternative by providing faster, approximated solutions that mitigate these disadvantages. While being computationally efficient, surrogate models also offer other advantages like differentiability, which is crucial for optimization tasks. However, the effective processing of FEA simulation data for training data-driven surrogate models remains challenging due to their complexity and size. Existing methods often rely on low-order data such as depth images exported from FEA simulations, limiting the surrogate models’ area of application. In contrast, data like 3D point clouds extracted from the FEA mesh, are more general and extensible to complex areas. Additionally, they present opportunities for bridging the Sim2Real gap by fine-tuning the model using transfer learning with point cloud data obtained from physical sensors. This paper introduces a novel approach using 3D point cloud data from FEA forming simulations to train surrogate models. In this paper we demonstrate the effectiveness of this method based on a two stage sheet metal forming process involving deep-drawing and trimming of a box-shaped component, with the goal of predicting the springback. Using machine learning architectures such as PointNet++ and Dynamic Graph Convolutional Neural Networks, we demonstrate that 3D geometric representations not only capture problem complexity more generally than 2D images but also achieve the same results as 2D state-of-the-art surrogate models.