Abstract
During the design cycle of safety-critical vehicle components such as B-pillars, crashworthiness performance is a key metric for passenger protection assessment. Traditional finite element (FE) simulations involve complex modeling and high computational demand. While machine learning surrogate models offer rapid predictions, they often struggle with complex 3D geometries and temporal stability. This paper proposes Recurrent Graph U-Net (ReGUNet), a new graph-based surrogate model for crashworthiness analysis. ReGUNet adopts a U-Net architecture with multiple graph downsampling and upsampling layers to improve efficiency and accuracy, while a recurrent mechanism enhances the stability of temporal predictions. Evaluated on a B-pillar side-crash case study, ReGUNet achieved a relative error of only 0.74% for maximum intrusion, reducing average prediction error by over 51% compared to baseline models with significantly improved computational efficiency.
Introduction
In automotive engineering, crashworthiness—the ability of a vehicle to protect its occupants during a collision—is a non-negotiable safety metric. Traditionally, engineers rely on high-fidelity Finite Element (FE) simulations to test designs. While accurate, these simulations are computationally expensive, often taking hours or days to run. This creates a bottleneck during the design phase, where engineers must iterate through hundreds of geometric variations.
Machine learning surrogate models aim to provide near-instant predictions by "learning" the underlying physics. However, standard neural networks struggle with the irregular, non-Euclidean geometry of 3D vehicle parts. While Graph Neural Networks (GNNs) are better suited for this data, they face two primary challenges: efficiency when processing thousands of nodes in a fine mesh, and temporal instability, where small errors accumulate over time steps, leading to physically unrealistic "rollout errors."
Methodology: The ReGUNet Architecture
ReGUNet addresses these challenges through three main innovations:
1. Graph U-Net Architecture
Similar to U-Nets used in image processing, ReGUNet utilizes hierarchical downsampling and upsampling layers. By creating "coarsened" versions of the mesh, information can travel across the entire component much faster. This allows the model to capture long-range physical dependencies—such as how an impact at the base of a pillar affects its top—without requiring hundreds of expensive message-passing steps.
2. Recurrence (Re-MP)
The model incorporates a recurrent mechanism (similar to a Gated Recurrent Unit) that propagates a "hidden state" across time steps. This acts as a temporal memory, ensuring that the prediction at step t+1 is physically consistent with step t, significantly smoothing the resulting deformation and preventing digital noise or "waviness."
3. Hybrid Downsampling
The model uses a two-tier approach: standard message passing at the finest level to handle varying mesh sizes, and "non-shareable weight" layers (DS-NW) at coarser levels. These DS-NW layers assign unique, trainable weights to specific edges, boosting accuracy for the shared structural features of the component.
Experimental Setup and Dataset
The model was trained and evaluated on a custom dataset of side-impact crash tests for a B-pillar component. Data was generated using high-fidelity FE simulations (PAM-CRASH), resulting in 200 samples (100 training, 50 validation, 50 testing). The geometry of the B-pillar was modified using mesh morphing and Latin Hypercube Sampling to ensure a diverse range of designs.
Each sample consists of a sequence of 12 graphs representing the component's deformation over 18 milliseconds. The use of multi-scale passing proved essential; by downsampling, the "receptive field" of the network grows, allowing it to understand the global structure of the part rather than just local clusters.
Key Results
ReGUNet demonstrates significant improvements over existing graph-based approaches:
- High Precision: The model predicted maximum B-pillar intrusion with a maximum relative error of only 0.74%.
- Superior Accuracy: ReGUNet reduced the mean prediction error by 51% compared to the best-performing baseline that lacked downsampling.
- Efficiency: The model used significantly less GPU memory (9.5 GB) compared to a non-downsampled recurrent baseline (20.3 GB), proving that hierarchical structures are essential for scaling GNNs to industrial-sized meshes.
- Stability: The "rollout error" was significantly lower and grew more slowly than in non-recurrent models, maintaining physical realism throughout the crash sequence.
Conclusion
ReGUNet represents a major step forward in applying deep learning to safety-critical automotive design. By combining hierarchical graph structures with temporal recurrence, the model achieves industrial-grade accuracy and efficiency. Future work will focus on extending the model to predict other vital metrics like crash force distribution and testing it on more complex, multi-part assemblies.
Summary updated