Large language model guided automated reaction pathway exploration

作者：Ruzhao Chen, Yang Liu, Zhe Chen, Yinwu Li, Fei Yang, Jiaxin Lin, Zhuofeng Ke · 发表于：Communications Chemistry · 年份：2025 · DOI：10.1038/s42004-025-01630-y · 被引用次数：5 · 研究领域：Machine Learning in Materials Science、Computational Drug Discovery Methods、Chemical Synthesis and Analysis

Fast and efficient automated exploration of reaction pathways is essential for studying reaction mechanisms and advancing data-driven approaches for reaction development and catalyst design. Here, we present a new program (utilizing Python and Fortran), capable of conducting automated, fast, and efficient exploration of reaction pathways for potential energy surfaces (PES) studies. This program integrates quantum mechanics and rule-based methodologies, underpinned by a Large Language Model-assisted chemical logic. Both active-learning methods in transition states sampling and parallel multi-step reaction searches with efficient filtering help enhance efficiency and accelerate PES searching. Its effectiveness and versatility in automating searches are exemplified through case studies of multi-step reactions, including the organic cycloaddition reaction, asymmetric Mannich-type reaction, and organometallic Pt-catalyzed reaction. ARplorer's capability to scale up for high-throughput screening significantly enhances its utility, positioning it as an efficient tool for data-driven reaction development and catalyst design.