Comparing traditional surrogate modelling and neural fields for vehicle crash simulation data
Résumé
In this study, we investigate the performance of classical surrogate modeling and neural fields for vehicle crash simulation, focusing on data that describe shell-like structures undergoing large deformations via a Lagrangian framework. While traditional surrogate modeling remains appealing for specialized physical simulations, particularly when data are limited, machine learning techniques such as neural fields offer a flexible alternative. However, our findings reveal that “off-the-shelf" neural networks may not adequately capture the complex physics inherent in large-deformation scenarios. To address this issue, we introduce an improved neural field architecture that embeds data insights into its design, leading to superior results. Our approach, evaluated against both ReCUR reduced order model and standard neural network strategies, demonstrates notable advantages in reproducing crash responses and highlights the potential for broad applicability in fields beyond solid dynamics.
| Origine | Fichiers produits par l'(les) auteur(s) |
|---|---|
| Licence |