Scholarlink

<p>Materials analysis involves procedures designed to determine the composition and properties of materials using measurement instruments, primarily aiming to uncover the mechanisms behind unexpected phenomena (e.g., discoloration) and support the development of effective countermeasures. To accelerate this analytical process, we herein introduce a retrieval-augmented conversational system that integrates a 6.6-billion-parameter language model—trained explicitly for materials analysis—with a database of over 26,000 historical cases related to the study of organic and inorganic materials, including quantum-beam analysis. The language model is developed from the ground up, optimized for domain-specific understanding and task-specific performance, and further refined through preference learning, incorporating partial edits from more than 40 domain experts. During interactions, the system generates hypotheses regarding potential underlying causes and recommends suitable analytical methods for testing these hypotheses. An integrated follow-up question mechanism proactively identifies and queries the missing information required to formulate optimal investigative strategies. Architectural optimizations, including a reduced-parameter encoder and grouped-query attention, enable on-premise inference using a single workstation equipped with only 16 GB of graphics processing unit memory. In evaluations simulating real-world use cases, the proposed system outperforms a size-matched general-purpose open-weight language model in over 70% of the cases, based on preference judgments regarding its helpfulness in enhancing the precision and speed of materials analysis. We present the design rationale and training methodology of the system, as well as representative examples of its conversational capabilities.</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

Development of a domain-tailored on-premise dialogue system for accelerated materials analysis

  • 期刊论文

  • 出版模式：

    开放获取

  • 发表日期：

    2026年6月23日

查看全文
  • Shigeaki Goto,
  • Michiaki Kamiyama,
  • Eiichi Sudo,
  • Hidehiko Kimura

摘要

Materials analysis involves procedures designed to determine the composition and properties of materials using measurement instruments, primarily aiming to uncover the mechanisms behind unexpected phenomena (e.g., discoloration) and support the development of effective countermeasures. To accelerate this analytical process, we herein introduce a retrieval-augmented conversational system that integrates a 6.6-billion-parameter language model—trained explicitly for materials analysis—with a database of over 26,000 historical cases related to the study of organic and inorganic materials, including quantum-beam analysis. The language model is developed from the ground up, optimized for domain-specific understanding and task-specific performance, and further refined through preference learning, incorporating partial edits from more than 40 domain experts. During interactions, the system generates hypotheses regarding potential underlying causes and recommends suitable analytical methods for testing these hypotheses. An integrated follow-up question mechanism proactively identifies and queries the missing information required to formulate optimal investigative strategies. Architectural optimizations, including a reduced-parameter encoder and grouped-query attention, enable on-premise inference using a single workstation equipped with only 16 GB of graphics processing unit memory. In evaluations simulating real-world use cases, the proposed system outperforms a size-matched general-purpose open-weight language model in over 70% of the cases, based on preference judgments regarding its helpfulness in enhancing the precision and speed of materials analysis. We present the design rationale and training methodology of the system, as well as representative examples of its conversational capabilities.