High-fidelity finite-element simulations provide accurate crashworthiness predictions, but their cost limits iterative design exploration.
Deep learning surrogates can reduce this cost, but many component-level models are developed under a single prescribed boundary condition, limiting generalisation to boundary variations.
This work proposes a Boundary-Condition-Aware Transformer Neural Operator (BAT-NO)
for autoregressive prediction of transient displacement fields and scalar crashworthiness responses under variations in geometry and boundary conditions.
A B-pillar simulation framework evaluates generalisation across variations in geometry, impact position and velocity, and support stiffness.
BAT-NO combines recurrent mesh processing with latent-grid Fourier operator processing.
Boundary-condition information is transferred to the latent grid through a hybrid local--global mechanism.
Slice-based attention models interactions among physically related regions, while direct boundary-to-grid projection preserves local spatial structure.
Validation Results
Across the validation sets for the shape-only, shape-and-loading, and shape-loading-boundary cases, BAT-NO achieves the lowest mean final-step mean nodal Euclidean displacement error among the evaluated baselines.
In the most challenging case, it reduces the mean error by 32.6% relative to the second-best model.
Hyperparameter tuning reduces the validation error from 0.451 to 0.269 mm, with a comparable error of 0.267 mm on 300 unseen test simulations sampled within the investigated design space.
Scalar Decoder and Crashworthiness Indicators
An attention-based scalar decoder jointly predicts six response trajectories with a mean relative error of 2.46%.
Most derived crashworthiness indicators have median errors below 3%.
These results show that explicit local and global boundary-condition representations improve crashworthiness prediction over expanded component-level design spaces.