Written by Sanghyuk Kim

Research

Scientific Machine Learning

This research develops a scientific machine learning framework for predicting 3D vehicle collision dynamics by decomposing the response into global rigid-body motion and local structural deformation. Full-scale crash simulations require high computational cost because they must resolve nonlinear contact, plastic deformation, and multi-scale spatio-temporal behavior. By embedding physical decomposition into coordinate-based neural networks, the framework learns continuous collision fields with improved accuracy and physical consistency. This enables faster simulation-based design evaluation when repeated finite element analysis is impractical.

Reference:

🔗 Project Page — Live Demo Available

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

Optimization Algorithm

2025_NEUCOM_graphical_absract

Deep neural network training requires both fast convergence and strong generalization. This research develops a projected variable three-term conjugate gradient algorithm that improves the training path of neural networks by reducing convergence to sharp minima while maintaining the stability of higher-order optimization. The method supports more reliable model training across language modeling, image classification, and engineering prediction tasks, providing an optimization strategy for deep learning models used in complex engineering problems.

Reference: