IBM and NASA have released an open-source artificial intelligence model designed to help researchers analyze lunar observations for tasks including crater mapping, volcanic-feature analysis and the identification of areas that may contain ice.

The NASA-IBM Lunar Foundation Model, announced on September 10, brings together observations collected at different resolutions and under sharply varying lighting conditions. IBM said the model is intended to give scientists a reusable starting point for lunar remote-sensing work instead of requiring a separate system to be trained from scratch for every task.

A model built for mixed lunar data

Lunar research draws on instruments that measure the Moon in different ways and at very different scales. NASA’s Gravity Recovery and Interior Laboratory mapped variations in the Moon’s gravitational field, while the Lunar Reconnaissance Orbiter has produced higher-resolution imagery and measurements used in studies of terrain and polar ice.

Reuters reported that the system was trained on more than 30 layers of data collected by nine instruments on four NASA missions. The accompanying model card describes a vision-transformer encoder-decoder trained on about 2 million co-registered lunar tile bundles spanning 11 modalities at two principal spatial scales. Some inputs represent about one meter per pixel, while others represent about 100 meters per pixel.

The system adapts the TerraMind approach previously developed by IBM and the European Space Agency for Earth observation. The lunar version incorporates information about acquisition geometry, including illumination angles, because the appearance of the Moon’s surface can change substantially with the position of the Sun. It also combines high- and lower-resolution samples within one training process.

The model card lists supporting data products from NASA instruments and Japan’s Kaguya, also known as SELENE, lunar orbiter. The model and its benchmark datasets are publicly available through Hugging Face under the Apache 2.0 license.

Initial tests cover ice, craters and volcanic features

IBM and NASA initially evaluated the model on three areas of lunar science: ice prospectivity near the poles, crater detection and mapping irregular mare patches, which are volcanic features that may help scientists study the Moon’s geological history.

For the ice-prospectivity benchmark, the project team reported that the model reduced error by 22 percent compared with a SwinV2 transformer used as a baseline. At a resolution of 100 meters per pixel, it outperformed the same baseline in crater detection by nearly 19 percent while using half the training data. At meter-scale crater detection, the results were described as comparable rather than superior. The model also posted a smaller improvement in mapping irregular mare patches.

The release is part of IBM and NASA’s Prithvi family of open foundation models for geospatial, weather and other scientific applications.

The release is for research rather than mission certification

The project’s documentation places limits on how the results should be interpreted. The model card says generated fields are not calibrated predictions and should not replace instruments, stereo photogrammetry or geodetic systems. It has not been validated for operational decisions such as certifying a landing site or clearing a route for hazards.

The ice output estimates prospectivity from a knowledge-based reference map rather than measuring ice directly. The documentation also says the model does not maintain an absolute geodetic reference frame and that its highest-resolution training coverage is distributed across selected sites rather than being globally dense.

Researchers can adapt the encoder for detection, segmentation and regression tasks using the released checkpoint and code. IBM said the team used lightweight low-rank adapters for its initial fine-tuning work, leaving most of the base model’s weights unchanged.

Open tools could broaden lunar research

Potential lunar ice is closely watched because water could support future crews and may also provide oxygen or material for producing rocket fuel. Better crater maps could help scientists study surface history and narrow the areas that require more detailed investigation.

NASA’s Artemis program is working toward renewed crewed exploration of the Moon, with a return currently planned for 2028, according to Reuters. The schedule and mission plans may change as the program develops.

The open-source release gives research groups outside IBM and NASA access to the model, datasets and evaluation materials. Its practical value will depend on independent testing, task-specific fine-tuning and validation against new observations before it is used in scientific or operational workflows.

Indian spacetech startup Pixxel nets $25M Series A funding led by Radical Ventures