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

Researchers introduce EvoMD-LLM, a framework that adapts large language models to predict molecular dynamics by treating chemical reactions as temporal sequences with duration-aware tokens. The model achieves 66.14% accuracy on prediction tasks and demonstrates the ability to generate explanations for its predictions without explicit supervision, suggesting LLMs can effectively ground themselves in physical simulations through symbolic temporal modeling.

EvoMD-LLM represents a meaningful convergence between natural language processing and computational chemistry. By discretizing reactive molecular dynamics trajectories into event sequences where each token carries temporal information, the framework enables standard autoregressive LLMs to capture the compositional evolution of chemical systems. This approach addresses a fundamental limitation of LLMs—their difficulty modeling temporal structure in dynamic physical processes—through temporal scaffolding, which treats event duration as explicit linguistic tokens that reduce hallucinated or chemically invalid outputs.

The research builds on growing interest in applying deep learning to molecular simulation, but distinguishes itself by reformulating the problem as language modeling rather than pure sequence prediction. Traditional approaches treat trajectories as continuous numerical data, while EvoMD-LLM's symbolic representation allows LLMs to leverage their strengths in compositional reasoning. The 66.14% accuracy significantly exceeds baseline neural networks, suggesting this paradigm captures meaningful patterns in molecular behavior.

The emergent capability to generate self-interpretations represents the most intriguing finding. Without paired trajectory-explanation training data, the model produces chemically grounded explanations for its predictions, implying LLMs can spontaneously integrate domain knowledge when problems are properly formulated. This has implications for scientific reproducibility and trust in AI-assisted molecular design.

For the AI and computational chemistry communities, this work demonstrates that symbolic temporal language modeling can effectively ground LLMs in physical simulations. Future applications could include drug discovery acceleration, materials science optimization, and reactive system prediction where interpretability matters. The framework's efficiency through fine-tuning rather than full retraining makes it accessible to researchers with limited compute resources.

• →EvoMD-LLM reformulates molecular dynamics as temporal language modeling, achieving 66.14% prediction accuracy by treating chemical reactions as discretized event sequences
• →Temporal scaffolding—representing event duration as explicit tokens—significantly reduces invalid or hallucinated molecular predictions compared to standard sequence models
• →The model generates chemically-grounded explanations for predictions without supervised explanation training, suggesting emergent interpretability in physically-grounded LLM applications
• →Framework outperforms sequential neural networks and language-based baselines on multiple temporal prediction tasks in reactive chemistry
• →Efficient fine-tuning approach makes the method accessible to researchers, with potential applications in drug discovery, materials science, and molecular design