How DeepMind learns physics simulators with Graph Networks (w/ author interview)

3-Second Blueprint (English)

1. Graph Network Simulators model fluids, sand, and deformables via local message passing to predict per-particle accelerations from data.
2. Permutation equivariance and shared edge-node weights yield high data efficiency by supervising every particle at each timestep.
3. Models trained on 2,500 particles generalize zero-shot to 28,000+ particles over thousands of steps, bounded by long-range forces.