Scholarlink

<p>This work presents an object-oriented software framework for automated data extraction and question-answering from materials science literature, with a focus on alloy properties. The system leverages Google’s Gemini API to process PDF documents, extract yield strength (YS), ultimate tensile strength (UTS), elongation, and additive manufacturing process data, and structure the results into tabular form. Extracted datasets are further enhanced with property cleaning, alloy composition parsing, and conversion to atomic percentages. A companion question-answering module allows users to query the resulting tables using natural language, supported by embeddings for similarity search. Performance benchmarks show reliable extraction accuracy, robustness to challenging figures, and scalable runtime behavior. Embedding analysis reveals that alloys cluster by processing route and display smooth gradients with mechanical properties, while cosine similarity between embeddings correlates negatively with differences in YS and UTS. As a case study, the extracted Al-alloy dataset was also used in supervised learning regressors for YS and UTS, achieving R<sup>2</sup> values of ~ 0.7–0.88, demonstrating that the mined data can directly support predictive modeling. The modular design and use of API-accessible LLMs lower the barrier to adoption, making this framework broadly accessible for materials informatics.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Enhancing Materials Data Workflows Through Object-Oriented Design and Large Language Models

  • 期刊论文

  • 发表日期：

    2025年11月14日

查看全文
  • Judah Immanuel,
  • Avik Mahata

摘要

This work presents an object-oriented software framework for automated data extraction and question-answering from materials science literature, with a focus on alloy properties. The system leverages Google’s Gemini API to process PDF documents, extract yield strength (YS), ultimate tensile strength (UTS), elongation, and additive manufacturing process data, and structure the results into tabular form. Extracted datasets are further enhanced with property cleaning, alloy composition parsing, and conversion to atomic percentages. A companion question-answering module allows users to query the resulting tables using natural language, supported by embeddings for similarity search. Performance benchmarks show reliable extraction accuracy, robustness to challenging figures, and scalable runtime behavior. Embedding analysis reveals that alloys cluster by processing route and display smooth gradients with mechanical properties, while cosine similarity between embeddings correlates negatively with differences in YS and UTS. As a case study, the extracted Al-alloy dataset was also used in supervised learning regressors for YS and UTS, achieving R2 values of ~ 0.7–0.88, demonstrating that the mined data can directly support predictive modeling. The modular design and use of API-accessible LLMs lower the barrier to adoption, making this framework broadly accessible for materials informatics.