Plox
NASA3 mins read

NASA and IBM Release Open-Source Lunar AI Model for Moon Science

NASA and IBM have released the Lunar Foundation Model, an open-source AI system trained mostly on 17 years of Lunar Reconnaissance Orbiter data to support lunar research, including polar ice prediction and crater detection.

The Decoder

The Decoder

NASA and IBM's open source lunar model turns 17 years of orbiter data into a foundation for lunar science

The NASA-IBM Lunar Foundation Model is designed to make decades of lunar observation data usable for machine learning, especially for polar ice prediction and crater detection.
Image credits:Nano Banana Pro prompted by THE DECODER

NASA and IBM Turn Moon Data Into an Open AI Foundation Model

NASA and IBM have released the NASA-IBM Lunar Foundation Model, described as one of the first open-source foundation models for lunar science. The model is meant to make decades of lunar observation data more usable for machine learning, especially in areas where labeled examples are limited. Its main value is flexibility: it can be adapted to specific research tasks instead of being built for only one narrow use case.

The Training Set Combines 17 Years of Lunar Observations

Diagram showing the NASA-IBM lunar model architecture, including lunar image data, elevation data, imaging geometry, tokenization, correlation learning, and downstream applications.
Image credits:NASA / IBM

The model was trained from scratch on SomBench, a co-registered multimodal lunar corpus containing nearly 2 million tile bundles across 11 modalities and two spatial scales. Most of the data comes from 17 years of Lunar Reconnaissance Orbiter observations, with additional data from GRAIL, Lunar Prospector, and JAXA’s Kaguya/SELENE probe. The full dataset brings together more than 30 spatially aligned data layers from nine instruments and four missions.

Lighting Geometry Is Treated as Core Context

The model is based on TerraMind, a multimodal Earth observation model, but it was trained from scratch for lunar data rather than fine-tuned. For each tile, it receives imaging geometry such as illumination angles, sun position, and tile extent. That matters because lighting conditions strongly affect how the Moon’s surface appears, so giving the model this context helps it interpret lunar imagery more directly.

Ice Prediction Shows the Clearest Performance Gain

Polar Moon maps comparing reference ice prospectivity data with predictions from ConvNeXt and the NASA-IBM Lunar Foundation Model.
Image credits:NASA / IBM

The model was tested on crater detection at 100-meter and 1-meter scales, polar ice deposit prediction, and Irregular Mare Patch segmentation. Its strongest reported result is in polar ice prediction, where it cut error by up to 22 percent compared with the best baseline it was tested against. It also beat SwinV2-B by nearly 19 percent on coarse-scale crater detection when trained with only half the data, according to IBM.

Useful for Analysis, Not a Replacement for Measurements

Examples of lunar crater detection comparing NASA-IBM model predictions with SwinV2-B predictions on WAC and NAC images.
Image credits:NASA / IBM

The researchers frame the model as a reusable foundation for downstream lunar science tasks, not a substitute for physical instruments or direct measurement. The article notes limits in absolute geodetic positioning, with latitude and longitude errors appearing in some generation tests. The model is publicly available through open-source channels, along with code, datasets, and benchmark collections, making it a practical starting point for researchers building lunar AI workflows.

Discover More