Fig. 3: Trajectories of the lattice active-matter model generated using the dynamics learned by the transformer.
From: Learning stochastic dynamics and predicting emergent behavior using transformers
The top row shows time-ordered snapshots of a trajectory generated at density \(\phi=0.12\overline{4}\), the value used during training. The two middle rows use densities ϕ = 0.3 and ϕ = 0.5; here, motility-induced phase separation can be seen. For comparison, the bottom row shows a trajectory generated with the original dynamics at ϕ = 0.5.
