NASA and IBM release open-source lunar AI model for exploration

YORKTOWN HEIGHTS — On Thursday, 10 September 2026, technology corporation IBM and NASA jointly announced the open-source release of the NASA-IBM Lunar Foundation Model, marking one of the first publicly accessible foundation models dedicated to the scientific exploration of the Moon. Creators position the release as a critical tool to help researchers automate surface mapping, analyze volcanic history, and identify potential water-ice deposits to support the establishment of a sustained human presence on the Moon.

Unifying Petabytes of Fragmented Lunar Archives

For decades, international space agencies have continuously observed the lunar surface, generating petabytes of multi-instrument data stored across disparate formats and resolutions. Historically, researchers have had to manually sift through massive archives of maps and images or rely on low-resolution, task-specific machine learning models that are computationally intensive and prone to accuracy gaps. To overcome these structural bottlenecks, IBM and NASA scientists built SomBench, described as the first unified, machine-learning-ready lunar dataset of its kind. As reported by Unite.AI, the corpus aggregates more than 30 spatially aligned layers derived from nine instruments across four missions. These include NASA’s Lunar Reconnaissance Orbiter (LRO) and Gravity Recovery and Interior Laboratory (GRAIL) missions, alongside complementary data from the Japan Aerospace Exploration Agency's (JAXA) SELENE/Kaguya mission. The SomBench dataset is organized into two primary tracks hosted on AWS under a CC BY 4.0 license: the Wide Angle Camera Low Resolution (WACLowRes) track containing 963,609 tiles covering 51.2 kilometers at 100 meters per pixel, and the Narrow Angle Camera High Resolution (NACHighRes) track comprising 1,000,113 tiles covering 512 meters at 1 meter per pixel. Splits are assigned at the Lunar Transverse Mercator zone level to prevent spatial data leakage.

Performance Metrics Across Lunar Phenomena

The foundation model itself uses a Vision Transformer (ViT-B) encoder-decoder architecture with 768 dimensions, 12 layers, and 12 attention heads. Trained across 16 H100 GPUs for 150,000 steps, the system incorporates acquisition geometry tokenization — such as illumination angles and solar-frame anchors, recognizing that lunar surface appearance is heavily dictated by lighting conditions rather than intrinsic material variation alone. Creator benchmarks highlight varying degrees of improvement over established baselines, such as the SwinV2-B model trained on ImageNet.

Lunar Analysis Task	Model Performance Metric	Baseline Comparison (SwinV2-B)
Potential Ice Deposits	RMSE reduction of up to 22–23%	RMSE of 0.0293 vs. 0.0377 (baseline)
Irregular Mare Patches (Volcanism)	3% improvement in capturing extent using imperfect labels	IoU of 0.5709 vs. 0.5687 (ConvNeXtV2-B)
Meter-Scale Crater Detection	Comparable accuracy with reduced computing overhead	Parity with SwinV2-B on WAC/NAC resolution
Context-Scale Crater Detection (~100m)	Outperforms baseline by nearly 19% using half the training data	Superior efficiency on wide-area feature extraction

Beyond raw accuracy, proponents emphasize resource efficiency. According to technical documentation, the model requires half the training data of comparable architectures while delivering higher resolution outputs at scale, making advanced celestial analysis more accessible to research groups with limited computing budgets.

Practical Applications and Commercial Integration

The open-source release feeds directly into broader institutional goals. Reuters reported that the tool supports the Artemis program's framework for a sustained human return to the Moon, scheduled for later this decade, by automating the identification of critical surface resources. Promotional materials emphasize that locating subsurface water-ice within permanently shadowed polar craters can supply astronauts with drinking water, breathable oxygen, and hydrogen-based rocket propellants necessary for deep-space missions. Furthermore, automated crater and boulder mapping assists mission planners in evaluating terrain age, selecting safe landing corridors, and mapping prospective locations for long-term lunar infrastructure. The model is distributed openly on Hugging Face under an Apache-2.0 license, with fine-tuning code maintained in a NASA-IMPACT GitHub repository and downstream adaptation executed via TerraTorch. It joins IBM's growing Prithvi family of open geospatial, weather, and heliophysics foundation models.

Frequently Asked Questions

Where can researchers access the new lunar foundation model and its dataset?

The NASA-IBM Lunar Foundation Model is published on Hugging Face under an Apache-2.0 license, while the accompanying SomBench dataset is hosted on AWS under a CC BY 4.0 license.

Related YouTube video

What specific missions contributed data to the training corpus?

The SomBench dataset integrates data from nine instruments across four missions, prominently featuring NASA's Lunar Reconnaissance Orbiter and GRAIL missions, alongside JAXA's SELENE/Kaguya mission.

Despite the enthusiasm surrounding its public release, the model card published by its creators establishes firm operational boundaries. Technical disclosures caution that the tool is not a scientific-grade generative product, maintains tens of meters of absolute geolocation uncertainty, and preserves no rigorous geodetic reference frame. Most notably, mission planners cannot rely on its outputs for active landing-site certification or hazard clearance. the model's ice-prospectivity predictions represent a knowledge-driven fuzzy-overlay map rather than direct physical confirmation of subsurface water-ice deposits, leaving researchers to balance open-source accessibility against strict operational validation requirements as upcoming lunar missions approach.