LLMs Show Promise for Mining Lunar Resource Data from Scientific Literature

Planning sustainable lunar missions hinges on understanding what materials are available on the Moon's surface. Yet the necessary compositional data exists scattered across countless scientific publications, making it labor-intensive to compile. Researchers have begun investigating whether large language models can accelerate this knowledge extraction process.

Applying LLMs to Lunar Science#

The approach capitalizes on LLMs' ability to process and synthesize information from technical documents at scale. For lunar mission planning and in-situ resource utilization (ISRU)—the practice of using local materials to support operations—rapid access to reliable composition data is essential. By automating the extraction of mineralogical and chemical information from published research, LLMs could significantly streamline the data collection phase that typically precedes mission design.

Current Capabilities and Limitations#

Testing off-the-shelf LLMs on this task reveals a mixed picture. The models perform reasonably well when extracting data presented in tabular formats, a common feature of scientific papers. However, the work also exposes important gaps. The lunar samples studied are highly heterogeneous, and their properties depend sensitively on subtle compositional variations. This means accuracy and uncertainty quantification become critical—errors in material characterization can have downstream consequences for mission feasibility.

Where Improvement Is Needed#

The research identifies two key areas for refinement. First, capturing detailed mineralogical information—the fine-grained description of mineral species and their distribution—remains challenging for current models. Second, performance degrades when handling more nuanced or complex information embedded in scientific text, rather than neatly formatted tables. These limitations suggest that while LLMs offer a promising starting point for automating lunar data extraction, the current generation of tools requires further development before they can fully replace expert human review.