Abstract
In recent years, various artificial intelligence-based surrogate models have been proposed to provide rapid manufacturability predictions of material forming processes. However, traditional AI-based surrogate models, typically built with scalar or image-based neural networks, are limited in their ability to capture complex 3D spatial relationships and to operate in a permutation-invariant manner. To overcome these issues, emerging graph-based surrogate models are developed using graph neural networks. This study developed a new graph neural network surrogate model named Recurrent U-Net-based Graph Neural Network (RUGNN). The RUGNN model can achieve accurate predictions of sheet material deformation fields across multiple forming timesteps. The RUGNN model incorporates Gated Recurrent Units (GRUs) to model temporal dynamics and a U-Net inspired graph-based downsample/upsample mechanism to handle spatial long-range dependencies. A novel 'node-to-surface' contact representation method was proposed, offering significant improvements in computational efficiency for large-scale contact interactions. The RUGNN model was validated using a cold forming case study and a more complex hot forming case study using aluminium alloys. Results demonstrate that the RUGNN model provides accurate deformation predictions closely matching ground truth FE simulations and outperforming several baseline GNN architectures. Model tuning was also performed to identify suitable hyperparameters, training strategies, and input feature representations. These results demonstrate that RUGNN is a reliable approach to support sheet material forming design by enabling accurate manufacturability predictions.
Introduction
The manufacturing industry faces a critical "speed vs. accuracy" trade-off in sheet metal stamping. Designing tools for components like car doors or structural bulkheads requires extreme precision; minor geometric errors can lead to defects such as wrinkling, cracking, or thinning. Traditionally, engineers rely on Finite Element (FE) simulations to predict these outcomes, but these methods are computationally intensive and slow, hindering iterative design processes.
While Machine Learning (ML) has been explored as a faster alternative, standard models like Convolutional Neural Networks (CNNs) are often restricted to 2D grid data. Real-world manufacturing involves complex 3D meshes that are better represented as graphs. Furthermore, standard Graph Neural Networks (GNNs) often struggle with long-range spatial dependencies and the high computational cost of modeling contact interactions between rigid tools and deformable sheets.
Methodology
The RUGNN framework addresses these challenges through an Encoder-Processor-Decoder architecture with several key innovations:
- Hierarchical U-Net Strategy: Inspired by the U-Net architecture in computer vision, RUGNN employs a graph coarsening mechanism. It reduces a fine mesh into hierarchical levels of coarser graphs, allowing message passing to efficiently capture long-range dependencies across the entire part before upsampling back to the original resolution.
- Temporal Memory with GRUs: To mitigate error accumulation over multi-step simulations, Gated Recurrent Units (GRUs) are integrated within the processor blocks. This allows the model to maintain material state information across timesteps.
- Node-to-Surface Contact Representation: Instead of computationally expensive contact edges, the model uses a lightweight representation. It calculates the distance and normal vector from each sheet node to the nearest tool surface, incorporating this information directly into node features.
Experimental Datasets
The model was evaluated using two distinct datasets generated via PAM-STAMP FE simulations:
- Dome-Shaped Cold Forming: A baseline study using an Al2008 alloy with 1,200 nodes. This case focuses on double-curvature surfaces, a standard test for material formability.
- Bulkhead-Shaped Hot Forming: A complex industrial-scale case involving 6,700 nodes and an AA6082 alloy. This study incorporates non-linear, temperature-dependent, and viscoplastic material behaviors under hot stamping conditions.
Results and Discussion
RUGNN demonstrated superior performance, achieving the lowest Mean Euclidean Error (MEE) compared to VanillaGNN, RGNN (Recurrent but no U-Net), and UGNN (U-Net but no Recurrence) baselines. The hierarchical structure not only improved accuracy by capturing global material stiffness but also provided significant memory savings by reducing the number of nodes during message passing.
The inclusion of GRUs proved essential for maintaining physical consistency over time, preventing the "error drift" observed in non-recurrent models. Furthermore, the model showed strong generalization capabilities, accurately predicting deformations for tool geometries not encountered during the training phase.
Conclusion
RUGNN represents a significant advancement in surrogate modeling for sheet metal forming. By effectively combining hierarchical spatial processing with temporal memory and efficient contact modeling, it provides FE-level accuracy at a fraction of the computational cost. Future research will focus on expanding the model to predict thinning fields and von Mises stress, integrating it into automated design optimization loops, and scaling the architecture to handle full-vehicle components with hundreds of thousands of nodes.
Summary updated