TL;DR
Get tech for your team delivered free — and shop member deals
- Fast, free delivery on millions of items
- Access to Prime Big Deal Days deals on October 6–7
- Prime Video, Amazon Music and more included
NASA and IBM have released an open-source lunar foundation model trained on nearly 2 million aligned data bundles, drawing mainly on 17 years of Lunar Reconnaissance Orbiter observations. The researchers report its strongest gains in predicting potential polar ice deposits, while results on some other tasks are comparable to existing models.
NASA and IBM Research have released an open-source lunar foundation model trained on nearly 2 million data bundles assembled from observations by four missions, including 17 years of Lunar Reconnaissance Orbiter data. The teams say the model can be adapted to research tasks such as mapping craters and predicting potential polar ice deposits, addressing a challenge in lunar science: abundant observations but relatively few labeled examples for training specialized algorithms.
The training corpus, called SomBench, combines data across 11 modalities and two spatial scales. It includes about 1 million high-resolution images from the Lunar Reconnaissance Orbiter’s Narrow Angle Camera, at roughly 1 meter per pixel, and just under 964,000 multispectral images from its Wide Angle Camera, at 100 meters per pixel. Data from NASA’s GRAIL and Lunar Prospector missions and Japan’s Kaguya/SELENE mission also contribute. The collection comprises more than 30 spatially aligned data layers from nine instruments and four missions, according to the report.
The researchers trained the model from scratch, rather than adapting an existing model to lunar data. It learns relationships among images, elevation information and imaging geometry by predicting portions of masked inputs. Because the Moon’s appearance changes substantially with illumination, the system receives details such as sun position and illumination angles as input. The dataset is divided geographically into map zones for training, validation and testing, rather than split randomly by individual image tiles, to limit overlap between those sets.
In tests covering polar ice prediction, crater detection at two scales and segmentation of Irregular Mare Patches, the technical report says the pretrained model matched or outperformed the comparison methods. IBM reports that the model reduced error on ice prediction by up to 22 percent compared with the SwinV2-B baseline. For coarse-scale crater detection, IBM reports a lead of nearly 19 percent; that result used half the data. On meter-scale crater detection and Irregular Mare Patch segmentation, however, results were broadly comparable to leading baselines, with differences within variation between training runs.
Making Lunar Data Easier to Use
The release could make it easier for researchers to build tools from the extensive but varied record of lunar observations. A pretrained model can be adapted to a particular task with fewer labeled examples than a model trained for that task from scratch, according to the report. That is relevant where scientists have large volumes of measurements but limited time or resources to label images and other data manually.
Potential ice deposits are of particular interest because permanently shadowed areas near the lunar poles can remain cold enough to preserve ice. If confirmed through further scientific work, such deposits could inform lunar research and future resource planning, including possible uses of water. The model’s predictions are not direct measurements of ice, however. They are a way to analyze existing observations and identify patterns for researchers to examine.
The mixed results also matter: reported gains are strongest for ice prediction and coarse-scale crater detection, not uniform across every task. The model’s open-source release may let researchers test it against other methods and adapt it, but its value will depend on independent evaluation and the quality of the underlying data for each application.
As an affiliate, we earn on qualifying purchases.
Four Missions, One Training Corpus
The model draws primarily on the Lunar Reconnaissance Orbiter (LRO), which has observed the Moon for 17 years. NASA says the volume of LRO data exceeds that of all other NASA planetary missions combined. The additional mission data broadens the training set beyond images: GRAIL contributes information about the Moon’s gravity field, Lunar Prospector data includes hydrogen measurements, and Kaguya/SELENE contributes mineralogical observations.
The project is described by NASA and IBM as one of the first open-source foundation models for lunar science. Unlike an algorithm designed for only one task, a foundation model is pretrained on a large collection of data and then adapted to specific uses. The researchers also trained on high-resolution and coarser imagery together. A technique called FlexiViT allows the same model to work with different image patch sizes without retraining, according to the report.
Evaluation included an architecturally identical model with random initialization as well as established comparison models. The researchers say even the randomly initialized control beat five of seven baselines on ice prediction, pointing to the model’s handling of multiple data types as one source of performance. This means the results cannot be attributed solely to lunar pretraining; the architecture and its separate processing paths for different data layers may also contribute.
Limits of the Reported Tests
The available source does not provide the full technical report, detailed test protocols or the numerical results for every task. The reported 22 percent reduction in ice-prediction error and nearly 19 percent improvement for coarse-scale crater detection are attributed to IBM; the precise evaluation conditions and comparison details are not fully set out in the source material.
It is also unclear how well the model will perform on observations or regions not represented in its training and test data, and whether independent researchers will reproduce the results. The authors report that differences on meter-scale crater detection and Irregular Mare Patch segmentation fall within variation between training runs, so those results do not establish a clear advantage. Predictions of possible ice deposits remain model outputs, not confirmation that ice is present at a location.
The source describes the model as open source but does not specify here the license, repository address or exact availability of all training data and code. Those details will affect how readily outside teams can inspect, reproduce and adapt the work.
Independent Testing and Lunar Data
The immediate next step is for researchers to examine the released model and compare its results with other approaches on relevant lunar datasets. Access to the code, model weights, documentation and data-use terms will shape how practical those checks are; the source material does not state those release details.
Further testing can clarify whether the reported gains hold across additional locations, instruments and lighting conditions, and whether they persist when labeled examples are scarce. For potential polar ice, model maps can help direct scientific attention, but any claims about actual deposits require evidence from lunar measurements and continued validation.
NASA and IBM’s announcement describes a research tool rather than a new lunar measurement or a confirmed discovery of ice. The model’s performance and usefulness will be clearer as outside researchers assess its methods and apply it to additional tasks.
Key Questions
What did NASA and IBM release?
They released an open-source lunar foundation model trained on observations from four missions. It is designed to be adapted for research tasks including crater detection, potential ice prediction and mapping volcanic features.
What data was used to train the model?
The SomBench corpus contains nearly 2 million tile bundles, mainly from 17 years of Lunar Reconnaissance Orbiter observations. It also includes data from GRAIL, Lunar Prospector and Japan’s Kaguya/SELENE mission.
Did the model confirm water ice on the Moon?
No. The model predicts patterns associated with potential ice deposits in polar regions. Those predictions are not direct measurements or confirmation that ice exists at a particular location.
Where did the model show its strongest reported results?
The strongest reported gains were in predicting potential polar ice deposits, where IBM says error fell by up to 22 percent against the SwinV2-B baseline. IBM also reports a nearly 19 percent advantage in coarse-scale crater detection, using half the data.
How did it perform on other tasks?
On meter-scale crater detection and Irregular Mare Patch segmentation, the model was broadly comparable to strong baselines. The report says differences on those tasks were within variation between training runs.
Source: rss
Fall Picks
fall essentials
As an affiliate, we earn on qualifying purchases.
