Materials Dual-Source Knowledge Retrieval-Augmented Generation for Local Large Language Models in Photocatalysts.

Area of Science:

• Materials Science
• Artificial Intelligence
• Scientific Research

Background:

• Large language models (LLMs) show promise as research assistants but struggle with domain-specific knowledge integration.
• Adapting LLMs for specialized fields like materials development requires incorporating both practical experimental data and theoretical insights.
• Ensuring data confidentiality necessitates offline operation, posing challenges for traditional LLM adaptation methods.

Purpose of the Study:

• To develop a retrieval-augmented generation (RAG) framework, MDSK-RAG, for specializing LLMs in materials development.
• To enable fully offline LLM operation for domain specialization, ensuring data confidentiality.
• To unify and leverage both experimental (CSV) and literature (PDF) data sources for enhanced LLM performance.

Main Methods:

• Developed the Materials Dual-Source Knowledge Retrieval-Augmented Generation (MDSK-RAG) framework.
• Converted tabular experimental data into text and retrieved relevant passages from both experimental and PDF literature sources.
• Summarized retrieved information using a local LLM and merged it with user queries before generation.

Main Results:

• The MDSK-RAG framework significantly improved the accuracy of local LLMs (e.g., gemma-2-9b-it) in materials science queries, increasing cosine similarity and expert ratings.
• Performance gains were statistically significant across multiple local LLM models, demonstrating the framework's effectiveness and scalability.
• The local LLM with MDSK-RAG outperformed a cloud-based LLM (GPT-4o) in a case study on metal-sulfide photocatalysts.

Conclusions:

• The MDSK-RAG framework provides a practical and extensible solution for adapting local LLMs to domain-specific scientific research, particularly in materials development.
• The framework successfully integrates offline experimental and theoretical knowledge, enhancing LLM accuracy while maintaining data confidentiality.
• Future work should address challenges in complex reasoning tasks where incomplete context retrieval can impact model performance.