Author(s)

Lewei Shi

Corresponding Author

Lewei Shi

Abstract

Solving for the dynamics of multi-body collision processes is a fundamentally difficult challenge in robotics, autonomous systems, and computer graphics. Classical physics simulators are computationally expensive, and only data-driven methods risk giving rise to violations of physical laws. We propose a new physics-informed graph neural network (PI-GNN) framework that leverages the representational ability of graph networks together with explicit physical principles for accurate and physically plausible trajectory prediction. Our model builds dynamic graphs, where objects are regarded as nodes and pairwise interactions are used to generate edges by multi-head attention modules. Through physics-based loss terms for energy conservation, momentum conservation, and collision constraints, the model attains 75.2% overall accuracy and achieves 71.8% physics-compliant performance across various scenarios. Experiments show up to 134%, 166%, and 154% improvements in energy, momentum, and overall efficiency over the baseline methods after simulator corrections. The model is able to generalize over different configurations, from simple two-body collisions up to complex five-body interactions, preserving physical coherence and reaching real-time inferential speed. Our model provides a stable basis for physics-based learning in dynamical systems.

Keywords

Graph Neural Networks, Physics-Informed Learning, Collision Prediction, Multi-Body Dynamics, Deep Learning