LunarLab 2025

Project Background

As international space agencies plan for a sustained human presence on the Moon, the ability to locate and effectively utilise in-situ resources becomes critical. Key lunar resources such as ice and oxygen are bound within compounds of titanium, iron and aluminium oxides — however these are unevenly distributed across the lunar surface and mostly inferred from disparate orbital sensor data. Furthermore, given that lunar observation is carried out by a vast number and range of instruments, several challenges arise when attempting to gain a comprehensive view.

The majority of recorded data is fragmented, as its sources are so numerous. A single lunar satellite may have various instruments including cameras, altimeters, spectrometers and radars — and there are many individual satellites that complicate the picture. The data ranges from optical and spectroscopic imagery to geophysical measurements such as gravity, and from thermal emissions to topography and reflectivity measurements. These datasets are typically stored in separate repositories in incompatible formats due to the differing data types and spatial resolutions.

The current best practice has resulted in the United States Geological Survey (USGS) Unified Geologic Map of the Moon — a product of time-consuming interpretation of datasets — but one that faces limitations due to human bias and cognitive restrictions. Some machine learning and deep learning approaches have yielded better models, but only in a single modality. The LunarLab team wanted to improve on these efforts and design a multi-modal foundation model combining the wide range of disparate data sources.

Project Results

The resulting foundation model, named Lunar-FM, is a significant step forward in planetary science — proving to be a scalable, self-supervised method for integrating 18 heterogeneous data layers across five modalities (optical, topography, thermal, radar, gravity) into a cohesive model for the first time. Lunar-FM is a unified, information-dense 768-dimensional latent embedding space that represents global lunar properties at a 0.5° × 0.5° resolution. The team's achievement standardises data representation, achieving a 300x data compression while retaining rich semantic information required for diverse downstream tasks including terrain classification, anomaly detection, rare feature discovery and global mapping based on sparse samples.

Due to the generalisable nature of a foundation model — that is, a single model can be adapted to many different downstream tasks — the team demonstrated this generalisability by using Lunar-FM to generate a global map of titanium dioxide (TiO₂) using only eight validated samples.

The agentic AI element, Lunar Analyst Copilot, bridges the gap between complex multidimensional lunar data and natural language queries by scientists, geologists and mission planners. The system utilises a Large Language Model (LLM) as its front-end, allowing users to interact with the vast lunar dataset using queries such as "where are the best places to land near the South Pole to search for water ice?". The LLM acts as an interpreter, translating the high-level human query into a specific sequence of executable data science tasks, understanding which underlying data modalities and models are required to synthesise an answer.

The capabilities of Lunar-FM support a range of critical applications for science, resource utilisation and mission planning, including:

• Identification of high-potential resource sites such as ilmenite-rich basalts, FeO-rich regolith or KREEP terrains rich in potassium, rare earth elements and phosphorus.
• Locating and characterising polar deposits, including mapping likely water-ice stability zones using thermal data.
• Selecting optimal regions of interest by combining topographical and physical constraints: slope, surface roughness, rock abundance and illumination conditions.
• Modelling thermal environments for lander, rover and habitat design, considering diurnal cycles, cold traps in permanently shadowed regions and overall thermal stress on hardware.
• Assessing regolith mechanical properties such as grain size and compaction to support excavation and in-situ resource utilisation (ISRU) operations.
• Detecting subsurface geological structures such as mass concentrations linked to mineral enrichment, primarily using gravity anomaly data.
• Estimating near-surface hydrogen abundance — a proxy for water content — for water extraction feasibility studies.
• Achieving a 20% performance boost in global titanium dioxide mapping using only ten expert-validated samples, illustrating how foundation models unlock planetary-scale insights with minimal overhead.

The potential to interact with the data via natural language democratises advanced analysis, providing easily operable tools for geologists and prospectors without requiring extensive coding expertise.

The Scan Partnership

Scan is a major supporter of LunarLab 2025 and FDL Europe, building on its participation in the previous five years' Earth Systems Lab events. As an NVIDIA Elite Solution Provider, Scan contributes multiple DGX supercomputers via Scan Cloud to facilitate much of the machine learning and deep learning development and training required during the research sprint.