GNN Surrogate Modeling for Crash Simulation
An applied ML research workflow for predicting mesh deformation, acceleration, and safety-relevant HIC15 from time-dependent CAE simulations.
Industrial AI / Graph Neural Networks / Scientific ML
An applied GNN surrogate-model research and engineering workflow for crash simulation prediction, evaluation, and validation.
Anonymized technical surrogate-model architecture overview
Overview
High-fidelity crash simulations generate large, graph-structured mesh datasets across multiple timesteps. This project investigated whether GNN-based surrogate models could provide faster approximate predictions of deformation and acceleration while preserving the engineering signals required for safety assessment.
My work focused on the applied ML and research workflow around the surrogate model rather than replacing the underlying physics solver. The project was executed as an industrial CAE collaboration with engineering software collaboration.
Why this was difficult
Imbalanced Node Dynamics
Most mesh nodes showed little deformation, allowing generic losses to be dominated by easy far-field regions while missing high-impact localized gradients.
Localized Safety Signals
Safety-relevant acceleration was concentrated in a small subset of nodes around the impact zone, making standard mean error metrics misleading.
Downstream Metric Sensitivity
Acceleration peaks strongly affected downstream HIC15 calculations, where minor trajectory smoothing caused catastrophic safety misclassification.
Autoregressive Error Accumulation
Multi-timestep rollout introduced error accumulation and initial-state mismatch as previous predictions were fed back into subsequent GNN steps.
Deceptive Aggregate Loss
A model could perform well on aggregate validation loss while still failing around the critical impact zone during engineering evaluation.
Multi-Objective Conflict
Different physical outputs required different evaluation and optimization strategies, as joint optimization of position and acceleration degraded peak accuracy.
My contribution
Data and workflow engineering
Built preprocessing and export workflows for graph-structured simulation data, covering mesh and metadata ingestion, feature attachment, dataset preparation, prediction remapping, and unified analytics outputs.
Model development
Designed spatial and historical features, custom graph-loss variants, and autoregressive rollout workflows for deformation and acceleration prediction using PyTorch and a commercial graph-learning toolkit.
Engineering evaluation
Developed evaluation logic beyond generic validation loss, including directional deformation errors, critical-region analysis, acceleration trajectories, HIC15 error, and safety-oriented diagnostics.
Experiment and diagnostic tooling
Automated repeatable experiment workflows and built interactive dashboards for comparing model trials, inspecting mesh-level errors, and analyzing deformation, acceleration, training behavior, and HIC outcomes.
System workflow
- 01Prepare simulation data
Parse mesh and metadata, validate snapshots, decimate elements, and attach model features.
- 02Build graph inputs
Construct node, edge, global, spatial, and historical features from CAE mesh topology.
- 03Train surrogate models
Run configurable GNN experiments and custom physics-aware loss variants.
- 04Roll predictions forward
Feed predicted deformation and acceleration back into subsequent timesteps autoregressively.
- 05Evaluate engineering outputs
Compare deformation, acceleration trajectories, critical regions, and HIC15 safety values.
- 06Diagnose and iterate
Use Streamlit dashboards and trial comparisons to isolate error propagation and failure patterns.
Technical decisions
Spatially aware features
spatial-feature experiment- Problem
- Prediction errors were concentrated around localized impact regions while far-field mesh nodes dominated the loss calculation.
- Decision
- Incorporated spatial features representing the topological and geometric relation between each mesh node and the impact location.
- Result
- Reduced combined acceleration and deformation MAE by approximately 15% in corresponding feature experiments.
Task-specific loss design
custom-loss experiment- Problem
- A standard uniform loss allowed the large number of non-deforming nodes to dominate training and weakened supervision of localized peak responses.
- Decision
- Separated deformation and acceleration supervision, applying higher loss weighting to localized high-response impact zones while maintaining lower weighting on far-field nodes.
- Result
- Reduced HIC prediction error by 27% in the evaluated custom-loss experiment.
Autoregressive timestep rollout
selected rollout experiment- Problem
- Treating each timestep independently failed to model temporal state transitions, while multi-step rollouts risked error propagation.
- Decision
- Implemented an autoregressive rollout workflow with recurrent state updates and targeted analysis of initial-state mismatch.
- Result
- In a selected rollout experiment, deformation MAE decreased by 13% and X-direction deformation R² improved from 0.75 to 0.86.
Results
These results come from separate experiments and model configurations. They should be interpreted individually, not as the performance of one combined final model.
- Baseline
- vs. baseline feature configuration
- Baseline
- vs. joint loss baseline
- Baseline
- vs. baseline rollout setup
- Baseline
- vs. baseline rollout setup
Anonymized Model Diagnostics
I built interactive Streamlit, Plotly, and PyVista diagnostics to replace manual comparison across simulation files, prediction exports, training logs, and engineering plots.
Streamlit / Plotly / PyVista. Conceptual reconstruction of the implemented analysis workflow.
- Truth vs. prediction acceleration curves
- Deformation error by direction and timestep
- Safety metric calibration and relative-error distribution
- Over-prediction vs. under-prediction analysis
- Trial and loss-function comparison
- 3D mesh-level error visualization using PyVista
- Impact-region and critical-deformation analysis
Lessons & takeaways
- Aggregate ML metrics can hide engineering-critical failure modes in localized impact zones.
- Different physical outputs require different loss and evaluation strategies rather than a naive joint loss.
- Rollout models must be evaluated for error propagation over time, not only single-step accuracy.
- Interactive 3D & curve diagnostics were essential for connecting ML results with engineering interpretation.