Abstract
Large Language Models (LLMs) have made significant advancements in natural language processing and human-like response generation. However, training and fine-tuning an LLM to fit the strict requirements in the scope of academic research, such as geoscience, still requires significant computational resources and human expert alignment to ensure the quality and reliability of the generated content. The challenges highlight the need for a more flexible and reliable LLM workflow to meet domain-specific analysis needs. This study proposes an LLM-driven workflow that addresses the challenges of utilizing LLMs in geoscience data analysis. The work was built upon the open data API (application programming interface) of Mindat, one of the largest databases in mineralogy. We designed and developed an open-source LLM-driven workflow that processes natural language requests and automatically utilizes the Mindat API, mineral co-occurrence network analysis, and locality distribution heat map visualization to conduct geoscience data analysis tasks. Using prompt engineering techniques, we developed a supervisor-based agentic framework that enables LLM agents to not only interpret context information but also autonomously addressing complex geoscience analysis tasks, bridging the gap between automated workflows and human expertise. This agentic design emphasizes autonomy, allowing the workflow to adapt seamlessly to future advancements in LLM capabilities without requiring additional fine-tuning or domain-specific embedding. By providing the comprehensive context of the task in the workflow and the professional tool, we ensure the quality of LLM-generated content without the need to embed geoscience knowledge into LLMs through fine-tuning or human alignment. Our approach integrates LLMs into geoscience data analysis, addressing the need for specialized tools while reducing the learning curve through LLM-driven interactions between users and APIs. This streamlined workflow enhances the efficiency of exploratory data analysis, as demonstrated by the several use cases presented. In our future work we will explore the scalability of this workflow through the integration of additional agents and diverse geoscience data sources.