Graph neural network-based dynamics simulation of aerospace mechanical bodies

In recent years, the application of Graph Neural Network (GNN)-based methods to simulate complex physical systems has brought new momentum to the fields of computational science and aerospace engineering. In spacecraft dynamics simulation, where accurate modeling of rigid body motion and complex force interactions is vital, GNNs offer a promising alternative to traditional physics-based methods. However, existing GNN-based approaches remain limited to relatively simple and idealized scenarios, restricting their applicability in real-world aerospace systems characterized by highly interconnected mechanical components and intricate force couplings. A major bottleneck lies in the difficulty of acquiring force-related data within spacecraft or similar industrial systems, which hinders the model’s ability to reason about contact, friction, and internal forces. As a result, current models often rely only on positional and velocity information, lacking robustness in handling force-dominant behaviors essential for accurate simulation. To tackle this challenge, we propose a novel constraint-guided GNN-based simulation framework tailored for spacecraft dynamics. Our method integrates explicit force analysis into the prediction process, modeling both contact and non-contact forces. Furthermore, we incorporate physical constraints derived from Kane’s equations to ensure that the learned dynamics adhere to real-world physical laws. Through rigorous theoretical formulation and extensive empirical validation, we demonstrate that our approach significantly enhances the physical fidelity and generalization ability of GNN-based rigid body dynamics simulators, making them more suitable for the complex and demanding requirements of spacecraft simulation.