从科学表格中自动提取材料知识，构建大规模数据库。
MatSKRAFT: A framework for large-scale materials knowledge extraction from scientific tables
- 将表格转为图结构，用约束驱动的图神经网络提取知识
- 属性提取F1达89.33，成分提取71.35，速度比现有模型快6-496倍
- 适用于材料发现、数据驱动研发的科研人员
科学进步越来越依赖于从海量文献中整合知识，但大多数实验数据仍困于半结构化格式，难以系统提取与分析。本文提出MatSKRAFT，一种可大规模自动提取并整合材料科学知识的计算框架。该方法将表格转化为基于图的表示，通过约束驱动的图神经网络（GNN）直接编码科学原理到模型架构中。在性能上显著超越当前前沿大语言模型：属性提取F1达89.33，成分提取F1达71.35；处理速度比最快模型快6倍，比最慢模型快496倍，且硬件需求低。应用于超过4.55万篇论文中的66,267张表格，构建包含509,281条记录的综合数据库，其中104,000种成分扩展了主流数据库覆盖范围。该系统揭示了以往被忽略的具有独特性质组合的材料，并支持数据驱动发现成分-性质关系，为材料与科学发现奠定基础。
原文摘要 · Abstract (English)
Scientific progress increasingly depends on synthesizing knowledge across vast literature, yet most experimental data remains trapped in semi-structured formats that resist systematic extraction and analysis. Here, we present MatSKRAFT, a computational framework that automatically extracts and integrates materials science knowledge from tabular data at unprecedented scale. Our approach transforms tables into graph-based representations processed by constraint-driven GNNs that encode scientific principles directly into model architecture. MatSKRAFT significantly outperforms contemporary frontier large language models, achieving F1 scores of 89.33 for property extraction and 71.35 for composition extraction, while processing data 6-496 times faster compared to the fastest and the slowest models respectively, with modest hardware requirements. Applied to 66,267 tables from more than 45,500 research publications, we construct a comprehensive database containing 509,281 entries, including 104,000 compositions that expand coverage beyond major existing databases. This systematic approach reveals previously overlooked materials with distinct property combinations and enables data-driven discovery of composition-property relationships forming the cornerstone of materials and scientific discovery.
Thank you to arXiv for use of its open access interoperability. PaperDance 不是 arXiv 官方产品；中文卡片由大模型生成，请以原文为准。