Research map / Architectures and efficiency / Attention and transformers

AtomWorld-Mem: Memory-Restored World States for Long-Horizon Atomistic Evolution

NeurIPS 2026 · Architectures and efficiency › Attention and transformers · approach: State space and recurrent sequence models · 0 citations (OpenAlex)

Read the paperCheck your idea against 108,967 papers

AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Proposes AtomWorld-Mem, a memory-restored world model that prioritizes legal vacancy-mediated KMC events in Fe-Cu alloy evolution using short- and long-term memory, reporting large fixed-budget acceleration over Zacros KMC.

Problem

Long-horizon atomistic evolution with kinetic Monte Carlo is slow because instantaneous local snapshots are ambiguous: similar configurations can hide different future event preferences and waiting-time scales.

Approach

Treat atomistic evolution as a partially observed world: keyframes from multi-scale spatial encoders are integrated by short-term event memory and long-term recurrent structural memory into a latent state that scores legal KMC events, with time still advanced by the simulator's residence-time rule.

Key results

• AtomWorld-Mem reaches 420.05x fixed-budget acceleration factor over Zacros at 0.18% Cu, versus 66.19x for snapshot-only AtomWorld.
• At 1.34% Cu, AtomWorld-Mem achieves 51.25x AF, versus 11.13x for snapshot-only and 2.16x for Energy-Greedy KMC.
• Zero-shot transfer across 16 unseen BCC alloy-temperature domains gives a median AUC gain of 22.9x and a maximum of 52.1x.

Surprising findings

• Full model AF at 0.18% Cu (418.44x in ablation table) is not clearly consistent with the 420.05x main-table value and the 100k-step energy plot described, suggesting possible reporting inconsistency.
• Energy-Greedy KMC gives only 2.16x AF at 1.34% Cu, far below the learned snapshot-only policy, indicating local energy descent is a weak heuristic.

Limitations

• Experiments are limited to vacancy-mediated BCC alloy evolution with a fixed event catalog.
• The biased policy does not preserve detailed balance or the unbiased KMC path measure.
• Temporal fidelity is only approximate; stronger clock-consistency objectives remain future work.

Experimental setup

Datasets: Fe-Cu alloy KMC simulations (Zacros)

Benchmarks: Fe-Cu KMC fixed-budget progress 16 unseen BCC alloy-temperature domains

Metrics: acceleration factor (AF) aggregation degree arrival hit DynErr TimeErr AUC acceleration ratio

Baselines: Zacros traditional KMC Energy-Greedy KMC Snapshot-only AtomWorld

Builds on: proximal policy optimization kinetic monte carlo

Closest papers by meaning

• All-atom Diffusion Transformers: Unified generative modelling of molecules and materials (ICML 2025 · similarity 0.84)
• FlashMD: long-stride, universal prediction of molecular dynamics (NeurIPS 2025 · similarity 0.84)
• CrystalREPA: Transferring Physical Priors from Universal MLIPs to Crystal Generative Models (NeurIPS 2026 · similarity 0.84)
• LLM Meets Diffusion: A Hybrid Framework for Crystal Material Generation (NeurIPS 2025 · similarity 0.83)
• EP-Flow: Disordered Crystal Structure Prediction without Site-level Annotation (NeurIPS 2026 · similarity 0.83)
• AutoMS: Multi-Agent Evolutionary Search for Cross-Physics Inverse Microstructure Design (ICML 2026 · similarity 0.83)