MomentumGNN: A Graph Neural Network for Deformable Objects
Researchers have developed MomentumGNN, a novel graph neural network architecture that accurately tracks linear and angular momentum in deformable objects. Unlike existing GNNs that predict unconstrained nodal accelerations, MomentumGNN predicts per-edge stretching and bending impulses, guaranteeing momentum preservation by construction. The model is trained unsupervised using a physics-based loss and outperforms baselines in scenarios where momentum is critical. This work addresses a key limitation of current GNNs in modeling dynamic behavior of deformable materials, offering improved physical accuracy for arbitrary shapes, mesh topologies, and material parameters.
Key facts
- MomentumGNN is a new architecture for graph neural networks.
- It predicts per-edge stretching and bending impulses.
- It guarantees preservation of linear and angular momentum.
- The model is trained in an unsupervised fashion.
- It uses a physics-based loss function.
- It outperforms baselines in momentum-critical scenarios.
- Existing GNNs struggle with momentum prediction.
- The work is published on arXiv.
Entities
Institutions
- arXiv