Generative Retrieval-Augmented Ontologic Graph and Multiagent Strategies for Interpretive Large Language Model-Based...

Area of Science:

• Materials Science
• Artificial Intelligence
• Computational Engineering

Background:

• Transformer neural networks, including large language models (LLMs), show potential in materials analysis, design, and manufacturing.
• LLMs can process diverse data types like language, symbols, code, and numerical data.

Purpose of the Study:

• To explore LLMs as tools supporting materials engineering analysis.
• To investigate LLM capabilities in information retrieval, hypothesis generation, and mechanistic relationship discovery.
• To develop and evaluate AI agent strategies using LLMs for materials analysis and design problems.

Main Methods:

• Utilized a fine-tuned LLM, MechGPT, trained on mechanics of materials data.
• Implemented retrieval-augmented Ontological Knowledge Graph strategies to address LLM limitations like hallucination and information recall.
• Employed agent-based modeling and nonlinear sampling strategies to enhance generative qualities.

Main Results:

• Fine-tuning improves LLM understanding of specific domains but does not prevent errors outside learned contexts.
• Retrieval-augmented knowledge graphs significantly enhance LLM generative performance and provide mechanistic insights.
• The relatedness feature in knowledge graphs offers advantages over standard retrieval augmentation, improving LLM performance and interpretability.

Conclusions:

• LLMs, augmented with knowledge graphs, offer powerful support for materials engineering tasks, including design and analysis.
• Knowledge graphs enhance LLM interpretability and allow for the integration of new data sources.
• Advanced strategies like agent-based modeling further improve LLM capabilities for complex generative tasks and active learning.