EvoMD-LLM: Learning the Language of Species Evolution in Reactive Molecular Dynamics

Researchers have proposed EvoMD-LLM, a framework for molecular dynamics that uses autoregressive large language models to learn compositional evolution over time, achieving up to 66.14% accuracy on temporal prediction tasks^[1].

EvoMD-LLM reformulates species-level molecular dynamics as a symbolic temporal language modeling problem, discretizing reactive MD trajectories into sequences of molecular events. The model outperforms sequential neural networks and language-based baselines, and can generate interpretations for its own predictions by incorporating relevant chemical knowledge^[1]. A related study on arxiv.org proposed an active learning framework that corrects machine-learned coarse-grained potentials at precise coverage gaps, achieving a 33.05% improvement in the Wasserstein-1 metric in Time-lagged Independent Component Analysis space^[2]. The active learning framework was submitted on September 21, 2025, and later revised, with the third submission on May 27, 2026, just a day before the submission of EvoMD-LLM on May 28, 2026. The proposed frameworks aim to improve the accuracy and efficiency of molecular dynamics simulations.