MMGUNet: AI surrogate cuts crash simulation costs with graph morphing

Nonlinear finite element crash simulations are the gold standard for vehicle safety design, but each run can take hours, making iterative optimization impractical. Graph neural networks (GNNs) promise faster surrogate models, but existing approaches struggle with generalizability: message-passing GNNs share weights across varying mesh graphs but lose accuracy, while edge-specific aggregation layers capture nonlinear relationships better but require fixed connectivity. The new Mask-Morph Graph U-Net (MMGUNet) solves this tension by morphing a fixed coarse graph hierarchy to match each input mesh's geometry using feature-aligned barycentric parameterization, then constructing cross-graph edges to preserve spatial correspondence. A further innovation is a two-stage training process: supervised pretraining with node masking, followed by freezing the high-parameter edge-specific layers and fine-tuning only the remaining parameters.

Evaluated across three challenging settings—in-distribution, out-of-distribution, and cross-component transfer (e.g., predicting on a different car part)—MMGUNet consistently outperformed both fixed-coarse-graph baselines and external methods like standard Graph U-Nets and message-passing GNNs. Key metrics were mean Euclidean distance (field accuracy) and maximum intrusion percentage error (critical for safety). The coarse-graph morphing alone boosted test accuracy significantly, and masked pretraining reduced train-test discrepancy while improving data efficiency during transfer. With 48 pages and 15 figures, the paper demonstrates a practical path toward reusable, data-efficient mesh-based surrogates for crashworthiness design exploration—potentially cutting months of simulation time in automotive R&D.