Automating structured data extraction from scientific literature is a critical challenge with broad implications across domains.We introduce nanoMINER,a multi-agent system combining large language models and multimoda...Automating structured data extraction from scientific literature is a critical challenge with broad implications across domains.We introduce nanoMINER,a multi-agent system combining large language models and multimodal analysis to extract essential information from scientific research articles on nanomaterials.This system processes documents end-to-end,utilizing tools such as YOLO for visual data extraction and GPT-4o for linking textual and visual information.At its core,the ReAct agent orchestrates specialized agents to ensure comprehensive data extraction.We demonstrate the efficacy of the system by automating the assembly of nanomaterial and nanozyme datasets previously manually curated by domain experts.NanoMINER achieves high precision in extracting nanomaterial properties like chemical formulas,crystal systems,and surface characteristics.For nanozymes,we obtain near-perfect precision(0.98)for kinetic parameters and essential features such as Cmin and Cmax.To benchmark the systemperformance,we also compare nanoMINER to several baseline LLMs,including the most recent multimodal GPT-4.1,and show consistently higher extraction precision and recall.Our approach is extensible to other domains of materials science and fields like biomedicine,advancing data-driven research methodologies and automated knowledge extraction.展开更多
基金supported by the Priority 2030 Federal Academic Leadership Program.Wewould like to thank S.Danilova,M.Reykina,and S.Shevtsova for their valuable contribution to the annotation of the dataset used in this study.
文摘Automating structured data extraction from scientific literature is a critical challenge with broad implications across domains.We introduce nanoMINER,a multi-agent system combining large language models and multimodal analysis to extract essential information from scientific research articles on nanomaterials.This system processes documents end-to-end,utilizing tools such as YOLO for visual data extraction and GPT-4o for linking textual and visual information.At its core,the ReAct agent orchestrates specialized agents to ensure comprehensive data extraction.We demonstrate the efficacy of the system by automating the assembly of nanomaterial and nanozyme datasets previously manually curated by domain experts.NanoMINER achieves high precision in extracting nanomaterial properties like chemical formulas,crystal systems,and surface characteristics.For nanozymes,we obtain near-perfect precision(0.98)for kinetic parameters and essential features such as Cmin and Cmax.To benchmark the systemperformance,we also compare nanoMINER to several baseline LLMs,including the most recent multimodal GPT-4.1,and show consistently higher extraction precision and recall.Our approach is extensible to other domains of materials science and fields like biomedicine,advancing data-driven research methodologies and automated knowledge extraction.
