High‑fidelity atomistic evolution over long timescales requires more than the current crystal configuration; it also needs the hidden dynamical context that instantaneous snapshots omit. Identical local atom arrangements can correspond to different future event preferences and waiting‑time scales, making a single snapshot ambiguous and turning long‑horizon evolution into a memory‑restored world‑state problem.
To address this, we introduce AtomWorld‑Mem, a memory‑restored atomistic world model. The evolving alloy is treated as an AtomWorld: spatial encoders write multi‑scale keyframes from dense local topology and sparse long‑range defect context. Short‑term event memory captures recent vacancy‑mediated jumps, while long‑term structural memory aggregates cross‑time structural cues; together they reconstruct a latent state that can predict future evolution.
The restored state is used to prioritize legal vacancy‑mediated events under single‑event Kinetic Monte Carlo (KMC) constraints. Event legality, physical execution, and residence‑time updates remain governed by the underlying simulator, ensuring that the original dynamics are untouched while the model steers the simulation toward more informative trajectories.
Empirically, AtomWorld‑Mem improves long‑horizon progress under a fixed microscopic event budget and preserves fidelity across energetic, structural, and vacancy‑transport observables. Crucially, it transfers zero‑shot to unseen alloy‑temperature AtomWorlds, indicating that the memory‑restoration mechanism captures reusable hidden‑state inference principles rather than a system‑specific local‑energy heuristic.
These findings position memory‑restored world‑state modeling as a promising route toward efficient, physically grounded, and transferable atomistic evolution.
Review: AtomWorld‑Mem’s explicit memory modules fill the gap left by instantaneous snapshots, offering an interpretable and transferable solution for long‑timescale atomistic simulations.