Do these 3 things before closing this tab:
1Clear out junk files and repair common Windows errors2Fix the driver behind crashes, sound loss and screen glitches3Repair Windows errors before they cause bigger problemsNASA and IBM released an open-source model for lunar remote sensing on September 10, 2026. It combines observations from multiple missions and instruments to help researchers map craters, study volcanic features, and estimate where conditions may favor polar ice. The Lunar Reconnaissance Orbiter’s long-running record is a major foundation, but the model’s training data are not simply 17 years of LRO observations: the release also draws on aligned data from other lunar missions.
What is the NASA-IBM Lunar Foundation Model?
It is a multimodal, multi-resolution foundation model designed to help scientists work with lunar observations that differ in instrument, viewing geometry, and spatial scale. NASA says its pretraining used high-resolution imagery and geophysical data from the Lunar Reconnaissance Orbiter (LRO), GRAIL, Lunar Prospector, and JAXA’s SELENE mission. The LRO’s extended observation record is an important part of the story, while the model itself combines aligned observations from several missions rather than treating one 17-year record as its entire training set. NASA Science
As an Amazon Associate I earn from qualifying purchases.
The related data assets are described in different ways by the project’s announcement and paper. IBM describes a companion dataset with more than 30 spatially aligned layers from nine instruments across four missions, including tens of thousands of images and maps. The authors’ 2026 paper describes SomBench as nearly two million geographically partitioned, co-registered lunar tile bundles across 11 modalities and two scales: 1 meter per pixel and 100 meters per pixel. Those figures refer to related but differently framed dataset descriptions. IBM’s announcement · The research paper
The model card identifies the architecture as a ViT-B encoder-decoder trained from scratch on SomBench. It conditions on acquisition geometry and jointly trains on meter- and hundred-meter-scale tiles. The released checkpoint is licensed under Apache-2.0, and the card documents use through TerraTorch. NASA-IBM model card
#1 Best Overall
What can it map or estimate on the Moon?
Craters and terrain
Crater mapping helps scientists characterize the Moon’s surface history and terrain. Such maps can support later analysis of landing hazards or surface infrastructure, but this model is not certified to make operational landing decisions.
Irregular mare patches and volcanic history
The model can support segmentation and mapping of irregular mare patches, unusual volcanic features relevant to the Moon’s thermal and volcanic history. Their ages and scientific interpretation remain questions for further study.
Rank #2
Polar ice prospectivity
The model estimates ice prospectivity using terrain, temperature, and other geophysical products. Its target is a knowledge-driven fuzzy-overlay map of conditions associated with possible ice, not a direct observation or measurement of water ice. Calling this “AI finding water” would overstate what the output establishes. NASA-IBM model card
How strong are the reported results?
NASA and IBM report task-specific benchmark comparisons, not one general accuracy score. The announcement summarizes comparisons with SwinV2-B ImageNet models; the percentages apply to the named tasks and settings, not to lunar science as a whole or to real-world mission performance. IBM’s benchmark summary
| Task and setting | Reported comparison |
| Polar ice prospectivity | Up to 22% lower RMSE than the SwinV2-B ImageNet model. |
| Irregular mare patch extent mapping | 3% better than the SwinV2-B ImageNet model in the cited comparison. |
| Crater results at approximately 100 m context scale | Nearly 19% better than SwinV2-B while using half the training data. |
| Meter-scale crater detection | Comparable accuracy to the cited state-of-the-art SwinV2-B comparator; the release describes greater efficiency and lower fine-tuning costs. |
The paper and model card report results across multiple seeds and show that relative performance changes with task and scale. At meter-scale crater detection, the best model and best baseline are nearly tied. A fair comparison with another approach should therefore match the task, resolution, training-data fraction, evaluation metric, adaptation strategy, and input coverage rather than comparing headline percentages in isolation. Research paper · Model card
What are the model’s limits?
- It is not a landing-site safety tool. The model card says the release has not been validated for landing-site certification or hazard clearance.
- Its ice output is a proxy. Prospectivity estimates do not measure or confirm ice.
- It may miss absolute location and values. The model card warns it can recover local structure while failing to preserve absolute values; generated latitude and longitude may be substantially wrong.
- Generated fields are qualitative. Multimodal generation is described as a qualitative probe, not calibrated scientific output.
- Its evaluated scope is lunar. It has not been evaluated beyond the Moon or on products absent from SomBench.
How can researchers access and use it?
NASA says the model is openly released on Hugging Face. Researchers can use the model card to access the checkpoint, review the Apache-2.0 license and documented limitations, and follow its TerraTorch fine-tuning workflow. The card recommends LoRA as a sensible default across many evaluated tasks; full fine-tuning performed best on the ice-prospectivity benchmark, while frozen-encoder results varied by task. These are findings from the reported research, not a guarantee that one strategy will suit every dataset. NASA Science · Model card
Rank #4
The authors released a pretrained checkpoint, benchmark datasets, and fine-tuning code. That makes the project a reusable starting point for research, not a turnkey mission system. As NASA chief science data officer and acting chief data and AI officer Kevin Murphy put it: “We also have to make data easier for scientists to explore and use.” IBM announcement
Quick Recap
Best Value
Product prices and availability are accurate as of the date/time indicated and are subject to change. Any price and availability information displayed on Amazon at the time of purchase will apply.
