NASA and IBM’s Open-Source Lunar Model Turns 17 Years of Orbiter Data Into a Foundation for Lunar Science
NASA and IBM Research have released an open-source lunar foundation model trained on 17 years of observations from the Moon. The model brings together data from multiple instruments and missions, creating a reusable foundation for lunar science.
The NASA-IBM Lunar Foundation Model is designed to support tasks such as crater detection, ice deposit prediction and segmentation of Irregular Mare Patches. The researchers say the model performs particularly well when predicting potential ice deposits at the lunar poles.
Nearly 2 Million Data Bundles From Lunar Observations
The team trained the model from scratch using SomBench, described as the largest co-registered multimodal lunar corpus to date. The dataset contains nearly 2 million tile bundles across 11 modalities and two spatial scales.
About 1 million high-resolution images come from the Narrow Angle Camera (NAC), with roughly 1-metre-per-pixel resolution. Meanwhile, just under 964,000 multispectral images come from the Wide Angle Camera (WAC), with a resolution of 100 metres per pixel.
Most of the data comes from 17 years of observations by NASA’s Lunar Reconnaissance Orbiter (LRO). The dataset also includes information from the GRAIL mission, Lunar Prospector and JAXA’s Kaguya/SELENE probe.
Overall, the collection combines more than 30 spatially aligned data layers from nine instruments and four missions. The researchers divided the dataset geographically by map zones to prevent leakage between training, validation and test data.
Model Accounts for Lunar Lighting Conditions
The NASA-IBM model is based on TerraMind, a multimodal Earth observation model. However, researchers trained the lunar model from scratch rather than fine-tuning the existing system.
For each tile, the model receives imaging geometry as explicit context. This includes illumination angles, the Sun’s position and the extent of the tile.
This approach addresses a particular challenge in lunar observation. Lighting geometry can strongly affect how the Moon’s surface appears. Therefore, providing this information directly can help the model distinguish illumination effects from actual surface properties.
The model also processes high-resolution and coarse imagery together. As a result, it can learn both fine surface details and broader geological context. FlexiViT further allows the trained model to adapt to different image patch sizes without requiring retraining.
Ice Prediction Shows the Strongest Gains
The researchers evaluated the model on crater detection at 100-metre and 1-metre scales, polar ice prediction and segmentation of Irregular Mare Patches.
The pretrained model matched or outperformed common baselines and a control model with random initialisation across the tested tasks.
The strongest improvement appeared in predicting ice deposits. Permanently shadowed regions near the lunar poles remain cold enough to preserve water ice for long periods. Such deposits could potentially provide water, oxygen and rocket fuel.
According to IBM, the model reduced prediction error by up to 22% compared with the SwinV2-B baseline.
For coarse-scale crater detection, the model also performed strongly. IBM reported an improvement of nearly 19% over SwinV2-B while using only half the available training data. This result suggests that the model can perform effectively with fewer labelled examples.
Multimodal Data Processing Improves Performance
Part of the model’s performance appears to come from how it handles different data types. The system gives each data layer its own processing path rather than treating every input as a simple stack of channels.
Interestingly, even the randomly initialised control model outperformed five of seven baselines on ice prediction without lunar pretraining. The researchers therefore identify data handling as one factor behind the results.
On metre-scale crater detection and Irregular Mare Patch segmentation, however, the model was broadly comparable with the strongest baselines. IBM reported a 3% advantage over SwinV2-B for Irregular Mare Patches, although the results were relatively close.
The researchers also tested LoRA, a lighter fine-tuning approach that trains only a fraction of the model’s parameters. LoRA generally matched full fine-tuning and performed better for crater detection, while full fine-tuning held an advantage on the two smallest tasks.
A Tool for Analysis, Not Physical Measurement
The lunar foundation model is intended to help researchers connect observations from different instruments and identify patterns that may be difficult to detect separately.
However, it cannot replace physical measurement instruments. Research results show that the model is not suitable for absolute geodetic positioning. In some generation tests, latitude and longitude differed by dozens of degrees, while reconstructed elevation structures could have shifted absolute height values.
The researchers therefore view the system as a reusable foundation for downstream scientific tasks rather than a substitute for direct measurements.
Further controlled experiments are also needed to isolate the contribution of individual model innovations. In addition, some test datasets remain relatively small.
Open-Source Lunar Model Joins Growing AI Research
The NASA-IBM Lunar Foundation Model has been made publicly available, alongside its code, machine-learning-ready pretraining datasets and benchmark collections. The model is also integrated into the open-source TerraTorch toolkit.
The project forms part of the NASA-IBM AI for Science collaboration. The organisations have worked together on foundation models under a Space Act Agreement since early 2022.
Their earlier work included the Prithvi model, which was released in 2023 and trained on Landsat and Sentinel-2 imagery for applications including flood and wildfire mapping.
IBM later developed TerraMind with the European Space Agency and Forschungszentrum Jülich for Earth observation. The lunar model builds on this multimodal foundation while adapting the approach to lunar data.
The work adds to a growing effort to apply foundation models to scientific observation. By bringing together long-term lunar measurements across instruments and missions, the NASA-IBM model provides researchers with a common foundation for analysing the Moon and developing new scientific applications.
Source: stratnewsglobal.com
