Empowering Chemistry Experts with Large Language Models for Literature Interpretation in Single‐Atom Catalysis Toward Advanced Oxidation

Abstract

Abstract Large language models (LLMs) hold considerable promise for large‐scale data extraction from scientific literatures for catalyst design and practical optimization. Yet, turning such outputs into reliable, formalized chemical knowledge would heavily rely on domain expertise rather than end‐to‐end automation. Herein, we present a human‐in‐the‐loop workflow integrating LLM‐facilitated structured data extraction with iterative, expert‐guided curation and analysis. As a proof of concept, we take single‐atom catalysts (SACs) for advanced oxidation processes (AOPs) as an example, enabling efficient data extraction, rigorous curation, and statistically driven interpretation. Thus, we uncover the key correlations among metal types, coordination environments, reaction substances, and catalytic performance, providing deeper mechanism insights into SAC‐driven AOPs. In contrast to fully automated, end‐to‐end models, our approach relies on human‐driven optimization at multiple stages, and underscores human insight as central to understand LLM outputs. By introducing human‐driven prompt refinement, model comparison, and expert‐led analysis, our method ensures that human cognition remains central to interpreting LLM outputs and converting structured data into reliable scientific knowledge. Our work addresses the limitations inherent in fully automated, end‐to‐end methodologies and effectively bridges the gap between structured outputs and catalytically meaningful insights.