Abstract
Crashworthiness assessment is a critical aspect of automotive design,traditionally relying on high-fidelity finite element (FE) simulations that arecomputationally expensive and time-consuming. This work presents an exploratorycomparative study on developing machine learning-based surrogate models forefficient prediction of structural deformation in crash scenarios using theNVIDIA PhysicsNeMo framework. Given the limited prior work applying machinelearning to structural crash dynamics, the primary contribution lies indemonstrating the feasibility and engineering utility of the various modelingapproaches explored in this work. We investigate two state-of-the-art neuralnetwork architectures for modeling crash dynamics: MeshGraphNet, andTransolver. Additionally, we examine three strategies for modeling transientdynamics: time-conditional, the standard Autoregressive approach, and astability-enhanced Autoregressive scheme incorporating rollout-based training.The models are evaluated on a comprehensive Body-in-White (BIW) crash datasetcomprising 150 detailed FE simulations using LS-DYNA. The dataset represents astructurally rich vehicle assembly with over 200 components, including 38 keycomponents featuring variable thickness distributions to capture realisticmanufacturing variability. Each model utilizes the undeformed mesh geometry andcomponent characteristics as inputs to predict the spatiotemporal evolution ofthe deformed mesh during the crash sequence. Evaluation results show that themodels capture the overall deformation trends with reasonable fidelity,demonstrating the feasibility of applying machine learning to structural crashdynamics. Although not yet matching full FE accuracy, the models achieveorders-of-magnitude reductions in computational cost, enabling rapid designexploration and early-stage optimization in crashworthiness evaluation.