Glass-forming liquids are dynamically heterogeneous: some regions relax quickly while others remain frozen, and the spatial pattern of fast and slow regions changes with temperature. Predicting which particles will move — the propensity for motion — from a single static snapshot has been a central challenge. The difficulty is that dynamics involves time, and a snapshot is timeless.
The T-BOTAN framework trains graph neural networks on static particle configurations and demonstrates something unexpected: the networks predict dynamic heterogeneity at temperatures not seen during training. A model trained on high-temperature configurations correctly predicts the relaxation patterns at lower temperatures, where the dynamics are qualitatively different (more heterogeneous, more cooperative).
More striking: the static configurations encode the macroscopic temperature itself. The network can infer what temperature a configuration was sampled from, using only the particle positions. No velocities, no trajectories, no dynamical information — just where the particles sit.
The implication: static structure contains thermodynamic state information beyond local structural details. The positions of particles in a snapshot are not merely consistent with a temperature — they are diagnostic of it. The arrangement carries enough information to reconstruct the dynamical behavior that will unfold from it. The snapshot knows its future. Not perfectly, not deterministically, but statistically — the correlations in position space map to correlations in trajectory space, and the mapping generalizes across temperatures. Structure predicts dynamics because structure is dynamics, frozen at an instant.