Momentum-Conserving Graph Neural Networks for Deformable Objects

MomentumGNN (arXiv:2604.26097) introduces a physics-constrained GNN that predicts per-edge stretching and bending impulses rather than unconstrained nodal accelerations, guaranteeing conservation of linear and angular momentum by construction. Trained unsupervised via a physics-based loss, it outperforms existing GNN baselines on deformable object simulations where momentum fidelity is critical. Robotics, cloth simulation, and FEM-replacement pipelines should take note, as prior GNNs like MeshGraphNets accumulate momentum drift that corrupts long-horizon rollouts. This is a meaningful architectural constraint, not a post-hoc correction.