Scholarlink

<p>Knowledge reconstruction is required for both existing and new material discovery. However, this process faces significant challenges, especially in ensuring the completeness and logical consistency of extracted information. We present a generalized method that can be used to reconstruct knowledge from inorganic science literature regarding synthetic routes and properties. Using a one-prompt design, a comprehensive dataset was built with the GPT-4 model, which was subsequently used to fine-tune four large language models (LLMs), including LLaMA3-8B-instruct (Llama), Gemma-7B (Gemma), Phi3-mini-128k-instruct (Phi), GPT3.5-turbo-1106 (GPT). The fine-tuned models demonstrated a strong capability in reconstructing synthetic routes for selective catalytic reduction (SCR). The information extraction (IE) process achieved high performance on the task of material synthesis route extraction, with a precision of 0.928, a recall of 0.957, and an F1 score of 0.962. The fine-tuned models were then used for large-scale IE, extracting 48,925 entities from 2205 articles. Due to the JSON structure of the initial GPT-4-constructed dataset, the extracted data was stored in the same format, providing a valuable resource for further research. The four fine-tuned LLMs were transferred to five other domains, including lithium-ion batteries (Li-ion), hydrogen evolution reaction (HER), oxygen reduction reaction (ORR), oxygen evolution reaction (OER), and methanol steam reforming (MSR), to test the generalized performance of the fine-tuned LLMs. The results show that the fine-tuned LLMs have high transferability across domains and can be applied to a wide range of materials science tasks. Ultimately, a comprehensive knowledge graph was constructed by integrating multiple entities (nodes) including article metadata, catalysts, preparation methods, raw materials, synthesis steps, synthesis conditions and performance metrics. These nodes are interconnected through various relational edges (such as synthesis method, synthesis action, temperature, pressure, time, atmosphere, equipment, conversion, selectivity, activity) that represent explicit physicochemical dependencies, procedural parameters, and performance correlations. This structured framework enables the efficient exploration of intricate relationships within materials science literature, thereby facilitating the rapid identification of high-performance catalysts and the data-driven optimization of synthesis.</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

Extracting and reconstructing knowledge in materials science literature using large language models

  • 期刊论文

  • 出版模式：

    开放获取

  • 发表日期：

    2026年1月8日

查看全文
  • Shuyuan Li,
  • Shihao Wei,
  • Chenyu Huang,
  • Yunjiang Zhang,
  • Guizhen Zhang,
  • Shaorui Sun

摘要

Knowledge reconstruction is required for both existing and new material discovery. However, this process faces significant challenges, especially in ensuring the completeness and logical consistency of extracted information. We present a generalized method that can be used to reconstruct knowledge from inorganic science literature regarding synthetic routes and properties. Using a one-prompt design, a comprehensive dataset was built with the GPT-4 model, which was subsequently used to fine-tune four large language models (LLMs), including LLaMA3-8B-instruct (Llama), Gemma-7B (Gemma), Phi3-mini-128k-instruct (Phi), GPT3.5-turbo-1106 (GPT). The fine-tuned models demonstrated a strong capability in reconstructing synthetic routes for selective catalytic reduction (SCR). The information extraction (IE) process achieved high performance on the task of material synthesis route extraction, with a precision of 0.928, a recall of 0.957, and an F1 score of 0.962. The fine-tuned models were then used for large-scale IE, extracting 48,925 entities from 2205 articles. Due to the JSON structure of the initial GPT-4-constructed dataset, the extracted data was stored in the same format, providing a valuable resource for further research. The four fine-tuned LLMs were transferred to five other domains, including lithium-ion batteries (Li-ion), hydrogen evolution reaction (HER), oxygen reduction reaction (ORR), oxygen evolution reaction (OER), and methanol steam reforming (MSR), to test the generalized performance of the fine-tuned LLMs. The results show that the fine-tuned LLMs have high transferability across domains and can be applied to a wide range of materials science tasks. Ultimately, a comprehensive knowledge graph was constructed by integrating multiple entities (nodes) including article metadata, catalysts, preparation methods, raw materials, synthesis steps, synthesis conditions and performance metrics. These nodes are interconnected through various relational edges (such as synthesis method, synthesis action, temperature, pressure, time, atmosphere, equipment, conversion, selectivity, activity) that represent explicit physicochemical dependencies, procedural parameters, and performance correlations. This structured framework enables the efficient exploration of intricate relationships within materials science literature, thereby facilitating the rapid identification of high-performance catalysts and the data-driven optimization of synthesis.