Physics‑Informed GNN Breakthrough for Accurate, Real‑Time Multi‑Body Dynamics

Recent advances in AI have excelled in vision and language, yet modeling complex physical systems remains challenging. Real‑world problems such as granular material flow, molecular dynamics, human motion and mechanical simulations involve multi‑body dynamics that must obey fundamental physical laws like momentum conservation.

Traditional numerical simulators are accurate but computationally expensive, while data‑driven models often ignore physical constraints, leading to error accumulation and divergence over long horizons. Physics‑informed machine learning has emerged to address this gap, with graph neural networks (GNNs) being particularly suitable because nodes can represent particles or rigid bodies and edges capture interactions.

To combine the expressive power of GNNs with physical inductive bias, EPFL researchers propose DYNAMI‑CAL GraphNet. The architecture explicitly embeds linear and angular momentum conservation into the network by assigning a local orthogonal reference frame to each edge, ensuring equivariance to 3D rotations, invariance to translations, and antisymmetry under node swapping.

The model operates in three stages:

Graph Representation : Multi‑body systems are encoded as graphs where node features include position, linear velocity, angular velocity, mass, charge, etc., and edge features encode relative positions and interaction types.

Scalarization‑Vectorization : Node and edge vectors are projected onto the edge‑local frames, producing high‑dimensional scalar embeddings that respect the required symmetries. These scalars are combined with node features to form order‑invariant edge embeddings.

Spatiotemporal Message Passing : Decoded forces and torques from edge embeddings are aggregated on nodes, updating linear and angular velocities. An implicit Euler integrator then updates positions, constituting one message‑passing layer. Latent edge memories are retained for subsequent steps, enabling true spatiotemporal reasoning.

Four heterogeneous datasets were used to evaluate generality:

Granular 6‑DoF collision dataset: 60 spheres with 6‑DoF motion in a box, sampled at 10⁻⁴ s (simulation step 10⁻⁶ s).

Constrained N‑body dataset: extended from Kipf et al., adding rigid rods and hinges.

CMU Motion Capture dataset: human joint trajectories for walking, running, jumping.

Protein molecular dynamics dataset: apo‑adenylate kinase (AdK) trajectories from MDAnalysis.

Experimental results show that DYNAMI‑CAL GraphNet consistently outperforms baselines (GNS, GMN, EGNN, ClofNet) in both single‑step error and multi‑step rollout stability. On the granular benchmark it preserves all particles, tracks kinetic energy decay, and maintains momentum consistency over 500 steps, whereas GNS diverges and loses particles. On constrained N‑body tasks it achieves the lowest prediction error across all configurations, and its inductive bias proves crucial for generalization. On human motion capture it yields the smallest error among all methods, maintaining stable joint trajectories over long rollouts, while GMN quickly diverges. On protein dynamics it captures fine‑grained vibrations and large‑scale conformational changes better than competing models.

These findings demonstrate that embedding physical conservation laws directly into GNN architectures yields models that are both accurate and physically interpretable, with strong generalization to unseen topologies and long‑term predictions. The work also highlights the relevance of such physics‑aware models for embodied AI, where robots, autonomous vehicles and intelligent manufacturing systems require reliable world models that can predict physical dynamics in real time.

For further details, see the Nature Communications paper “A physics‑informed graph neural network conserving linear and angular momentum for dynamical systems” (https://www.nature.com/articles/s41467-025-67802-5).