Written by Sanghyuk Kim

Research

Scientific Machine Learning

Vehicle Collision Dynamics

This study presents a rigid-deformation decomposition AI framework for 3D spatio-temporal prediction of vehicle collision dynamics. Conventional implicit neural representations struggle to capture high-frequency deformation modes in crash simulations. We address this by decoupling global rigid-body motion from local structural deformation through two scale-specific networks (RigidNet and DeformationNet), achieving a 29.8% reduction in rigid-body motion error, a 17.2% reduction in total interpolation error, and a 46.6% reduction in angular extrapolation error compared to undecomposed baselines. The framework recovers 92% of the directional correlation and 96% of the spatial deformation localization accuracy relative to an oracle upper bound, enabling physically consistent predictions for nonlinear collision dynamics.

Reference:

  • Kim S, Seo M, Yang S, Kang N (2025) Rigid-Deformation Decomposition AI Framework for 3D Spatio-Temporal Prediction of Vehicle Collision Dynamics. arXiv preprint arXiv:2503.19712. https://arxiv.org/abs/2503.19712
🔗 Project Page — Live Demo Available

Try the live demo of vehicle collision simulation on the project page.

Optimization Algorithm

2025_NEUCOM_graphical_absract

This study introduces the Projected Variable Three-Term Conjugate Gradient (PVTTCG) algorithm, designed to overcome the trade-off between fast convergence and strong generalization in deep neural network training. By stabilizing the optimization path through geometric projection, PVTTCG consistently improves generalization performance across benchmarks and real-world applications such as vehicle crash prediction.

Reference: