A new AI model predicts how materials evolve over years of simulated time without replaying every atomic collision along the way.
Researchers built AtomWorld-Mirror, a macro-step world model that compresses short bursts of atomic micro-events into direct jumps between structurally significant states. Instead of simulating every individual atom hop, it jointly predicts sparse structural edits and the physical time those edits represent, while enforcing constraints like local reachability, inventory conservation, and continuous-time consistency so the shortcuts stay physically valid. The team tested it on five systems, including irradiation aging in Cu-rich reactor pressure vessel steel, a Cu-Zr metallic glass, and a Li3N-based solid electrolyte. Across those cases, macro-step inference ran 1,000 to 10,000 times faster than conventional event-by-event simulation.
That gap matters because atomistic simulators spend most of their compute budget on low-impact local shuffling before anything structurally interesting happens, which is the long-horizon bottleneck that keeps materials scientists from simulating years of real-world aging or diffusion. A model that skips straight to the next meaningful state could make it practical to study slow processes, like reactor steel embrittlement, that are currently too expensive to simulate at full resolution.
The speedup is a range, not a guarantee, and it's demonstrated on five benchmark systems chosen by the authors, not yet the messy, unpredictable materials engineers actually worry about in the field.