NASA and IBM Open-Source AI Model That Maps the Moon
By @sharedot · · 8 pages
- AI
- Space
- Lunar Science
- Open Source
NASA and IBM released the open-source Lunar Foundation Model, trained on 17 years of lunar data, which they report cut ice-mapping error by up to 22%.
What happened: a shared AI backbone for lunar science
The model was trained primarily on 17 years of Lunar Reconnaissance Orbiter observations—roughly 2 million image tiles—alongside data from NASA's GRAIL and Lunar Prospector missions and JAXA's SELENE/Kaguya spacecraft.
The release also includes a companion dataset combining more than 30 spatially aligned layers from nine instruments across four missions, which the organizations describe as the first open-source lunar dataset of its kind to unify multimodal, multiresolution observations in a machine-learning-ready framework.
Why it is surprising: training on knife-edged shadows
Training a vision model on lunar data required rethinking standard computer-vision practice.
According to Yahoo Tech, IBM's Dr. Juan Bernabé-Moreno described the Moon's shadows as "knife-edged and pitch black," carrying no information, and said that asking the model to reconstruct masked image regions was "a complete disaster" because thousands of craters look nearly identical from orbit.
Yahoo Tech reports IBM's solution was spatial partitioning: slicing the Moon into wedge-shaped sections and completely separating training wedges from testing wedges to give the model consistency.
The model also passed an unplanned real-world test—Yahoo Tech reports that after a SpaceX Falcon 9 rocket stage impacted the Moon on August 5, the model correctly identified the new crater on the first attempt, even though the impact closely overlapped an existing one.
The evidence: benchmarks against a strong baseline
The performance claims come from a technical paper authored by IBM and NASA researchers and should be read as the organizations' own reported results.
According to TechRepublic, the model reduced root mean square error by as much as 22% compared with the SwinV2-B image model when predicting areas with high potential for lunar ice, and improved identification of irregular mare patches' extent by 3% when working with imperfect labels.
TechRepublic also reports that at meter-scale resolution the model delivered crater-detection accuracy comparable to current specialized methods, while at roughly 100-meter context resolution it outperformed SwinV2-B by nearly 19% using half as much training data.
IBM and NASA also released benchmark collections so independent researchers can reproduce and stress-test the comparisons.
The caveat: prospectivity is not proof of ice
NeoTeo emphasizes the most important limitation: the model does not directly detect lunar ice.
It estimates prospectivity—areas that match conditions associated with a greater possibility of stable ice on or below the surface—so a value on a prospectivity scale is an analytical output, not a drill core or direct confirmation of frozen water.
NeoTeo also cautions that the task-specific comparisons against SwinV2-B vary by task and operating conditions and should not be converted into a blanket claim that the model beats every alternative at every lunar task, and that the system does not certify safe landing sites or clear operational hazards.
TechRepublic similarly notes the results come from the organizations behind the model, and independent researchers will need to test it across additional datasets before its broader performance becomes clear.
The stakes: resources, science and Artemis planning
The applications map directly onto NASA's return-to-the-Moon ambitions.
Lunar ice in permanently shadowed regions could provide water and oxygen for future bases and potentially rocket fuel for missions deeper into space, so knowing where ice may be stable shapes where astronauts live and whether in-situ resources become viable.
Crater mapping helps scientists date lunar terrain and reconstruct solar system history, while also helping mission planners identify slopes, boulders and other hazards as NASA prepares for future surface missions following the Artemis II crewed flight around the Moon.
The model also accelerates study of irregular mare patches—young volcanic features that challenge established timelines for lunar cooling—helping scientists piece together the Moon's thermal evolution.
What comes next: open science and independent testing
The model joins IBM's Prithvi family of open foundation models covering weather, geospatial analysis and heliophysics, extending a NASA-IBM collaboration that includes the Surya heliophysics model, according to phys.org.
Because everything is open—the model, datasets, benchmarks, companion paper and TerraTorch toolkit integration—researchers can inspect the system, reproduce its reported results and build specialized tools without assembling data pipelines from scratch.
phys.org reports the science team included experts from universities and institutes such as the SETI Institute, Howard University and NASA's Goddard and Ames centers.
The release could lower the technical and computing barriers to lunar-analysis tools, but its broader value now depends on whether independent researchers can reproduce the reported results and validate outputs against direct observations before such maps inform mission decisions.
Sources
- techrepublic.com › NASA and IBM Launch Open AI Model for Lunar Research
- phys.org › Release of open AI model trained on 17 years of lunar data maps ice, craters and volcanoes
- aithority.com › IBM and NASA Release Open-Source AI Model to Support Lunar Exploration
- tech.yahoo.com › NASA and IBM Release Open-Source AI to Map the Moon
- neoteo.com › NASA and IBM release an open AI model for lunar science