Towards Large Language Models for Lunar Mission Planning and In Situ Resource Utilization

TL;DR

The paper investigates using large language models to extract lunar composition data from scattered scientific literature to support mission planning. It proposes a preprocessing pipeline that converts PDFs of lunar publications into a structured interval dataset, using a prompting-based approach to summarize compositions from Apollo-era samples (LSC) and to generate ground-truth intervals. The results show that providing the source papers to the model substantially improves accuracy (midpoint errors typically within a few percent and relative errors under ~5%), outperforming standalone prompts, though challenges remain for fine-grained mineralogy and multi-document synthesis. The work demonstrates the viability of a tool-enabled dataExtraction workflow for lunar resource assessment and outlines concrete directions for enhancing preprocessing, uncertainty quantification, and integration into autonomous mission-planning agents.

Abstract

A key factor for lunar mission planning is the ability to assess the local availability of raw materials. However, many potentially relevant measurements are scattered across a variety of scientific publications. In this paper we consider the viability of obtaining lunar composition data by leveraging LLMs to rapidly process a corpus of scientific publications. While leveraging LLMs to obtain knowledge from scientific documents is not new, this particular application presents interesting challenges due to the heterogeneity of lunar samples and the nuances involved in their characterization. Accuracy and uncertainty quantification are particularly crucial since many materials properties can be sensitive to small variations in composition. Our findings indicate that off-the-shelf LLMs are generally effective at extracting data from tables commonly found in these documents. However, there remains opportunity to further refine the data we extract in this initial approach; in particular, to capture fine-grained mineralogy information and to improve performance on more subtle/complex pieces of information.

Figures (12)

• Figure 1: Envisioned tool-enabled LLM for mission planning. In this paper, we focus on the preprocessing task of extracting structured information from scientific documents (blue boxes) which can then serve as a source of auxiliary knowledge to support interactive mission planning (future work). The collection of tools listed here is non-exhaustive.
• Figure 2: Example of chemical composition data from LSC document 10047.pdf. Note these tables exhibit various complexities, including analyses by multiple authors, blank entries, multiple units (percent, ppm, ppb), and rows where the units are implied based upon entries above.
• Figure 3: Excerpt from the LSC document for sample 14321. Here, the columns denoting different studies also designate the different mineralogical phases (termed "clasts" in the column header) derived from this sample. Our current ground truthing and LLM prompting strategy does not attempt to disambigute among the various phases within a sample; however, this would be a natural next step and has implications for performance analysis. For example, the "granite" column demonstrates significant compositional differences relative to the others.
• Figure 4: Weight percentages for four relatively abundant oxides. Blue intervals denote the manually extracted ground truth, green values denote data extracted by ChatGPT4o when the paper content is included within its context window, and red denotes the baseline ChatGPT4o result when the paper content is not provided. The green and blue intervals demonstrate much closer alignment across the entire range of samples. When an interval is small relative to the y-axis, it is rendered as a an hourglass-shaped marker rectangle (e.g., the blue and green values SiO2 for sample 12057). Note the x-axis is the same for all subplots whereas the y-axis is not. Also note the lack of explicitly defined ground truth for SiO2 in samples 71576 and 71595; in these cases, the LLM have provided an answer (despite the table entry being blank, see \ref{['fig:lsctable71595']}).
• Figure 5: Excerpt from LSC document 71595.pdf. Note that SiO2 quantities are not explicitly reported for these samples (and, hence, not included in ground truth; see also \ref{['fig:per-material-intervals']}).