NASA and IBM's open lunar model trains one set of weights across a 100-fold resolution gap
NASA and IBM's open lunar model ingests 11 instrument modalities across a 20,000-to-1 resolution range. USRA's account of how those observations were aligned is the part that explains the result.
The Engineer · Build desk
What happened
- NASA and IBM released the NASA-IBM Lunar Foundation Model as an open model on September 10, built for crater mapping, volcanic-feature analysis and research into possible polar ice.
- Pretraining ran from scratch on nearly two million geographically partitioned data bundles spanning 11 modalities, including camera imagery, topography, radar, mineralogy, gravity and illumination geometry.
- USRA's September 18 account credits planetary scientist Rachel Slank with aligning observations from instruments running 1 meter to 20 kilometers per pixel while keeping their physical context.
- For the SomBench evaluation suite, Slank manually identified more than 49,000 craters in Narrow Angle Camera imagery and paired the images with co-registered terrain models.
- The pretrained model matched or beat comparison systems initialized with ImageNet weights and an identical architecture started from random weights, according to the September 8 technical paper.
Why it matters
- What doesn't add up contradiction The SomBench paper counts more than 30 co-registered layers from ten instruments across four missions; IBM's release counts nine instruments across the same missions. A team reusing the corpus has to settle which list the published weights actually saw.
- What it limits constraint Because the model consumes solar angles and tile coordinates directly, an adopter needs that geometry recorded alongside their own tiles. Imagery without it supplies the network less than it was trained on.
- What it costs cost At about 1,100 H100 GPU-hours, the training bill is the easy part to reproduce. The expense sits in cross-instrument alignment and expert annotation, and those hours are labour.
- What it sets up precedent Shipping a hand-checked crater benchmark next to the weights sets an expectation for the next domain model: on the hardest task, the ground truth is something someone verified by eye, and a held-out slice of the same scrape will not do.
Next story · Science · 23h ago
More than 10,000 young faults put the moon's shrinkage at about 45 metersConfirmed3 publishers
Shadows can dominate a lunar image even when the terrain underneath changes little [3]. The model therefore takes acquisition geometry as an explicit input, including solar angles and tile coordinates, all of it information already recorded with the observation [4]. Of everything in this release, I would copy that first.
The pretraining corpus comes in two pools: 963,609 Wide Angle Camera bundles at roughly 100 meters per pixel, and 1,000,113 Narrow Angle Camera bundles at roughly 1 meter per pixel [6][7]. Add them and the total is 1,963,722 bundles, with the meter-scale pool at 50.9 percent [18][19]. Both scales go through one set of weights [5]. Each bundle can carry surface imagery, terrain measurements and contextual data such as illumination [8]. The instrument suite Slank had to reconcile spans 20,000 to 1 in resolution [22].
Divide the model card's compute figure by eight GPUs and the run is 137.5 hours of wall clock on a single node, under six days [20]. Across the corpus that is about two GPU-seconds per bundle [21]. The scarce input was the label set. At meter scale, small differences in lighting, slope or image resolution can resemble geological boundaries [15], so somebody had to decide where each rim was, and hand annotation does not scale with GPU count. The person drawing them had used Lunar Reconnaissance Orbiter thermal-inertia data during her master's work to investigate possible lunar cavities and lava tubes [17].
The September 8 technical paper evaluates three task families: crater detection at regional and meter scales, segmentation of irregular mare patches, and regression of lunar polar ice prospectivity [9]. The useful control there is the identical architecture trained from random weights, since it separates what pretraining bought from what the architecture bought [10]. For the scores to say anything about another workload, the tiles have to come from the same instruments and be co-registered the same way.
runtimewire reports that the pretraining code is missing and that this limits full reproducibility [11]. That constraint bites harder here than it would on a frontier-scale training run. A group wanting to retrain with a different modality mix or a different tiling scheme is not blocked by hardware, because the published run fits inside a week on one eight-GPU node [20].
What to watch
- Whether IBM publishes the pretraining code, so an outside team can rerun the path from bundles to weights.
- Whether the model card's instrument list is reconciled with the SomBench paper's count of ten instruments.
- Meter-scale crater detection results reported on imagery from outside the Lunar Reconnaissance Orbiter Camera.
Clarity's read
What the record supports and how the coverage leans. The claims behind it follow.
Reality
- Evidence60
- Adoption15
- Hype gap+10
- Incentives55
- Confidence65
Claim ledger
Ranked by verification strength, evidence, and original report placement.
- [1]
Rachel Slank helped NASA and IBM turn decades of mismatched Moon observations into the NASA-IBM Lunar Foundation Model, an open model released on September 10th for crater mapping, volcanic-feature analysis and research into possible polar ice.
- [2]
The model was trained from scratch on nearly two million geographically partitioned data bundles spanning 11 modalities, from camera imagery and topography to radar, mineralogy, gravity and illumination geometry.
- [3]
Illumination is especially important on the Moon, where shadows can dominate an image even when the underlying terrain changes little.
- [4]
The model receives acquisition geometry as an explicit input, including solar angles and tile coordinates, rather than being asked to infer information already recorded with the observation.
- [5]
The model trains across both image scales using one set of weights, covering a 100-fold resolution gap.
- [6]
SomBench's pretraining corpus contains 963,609 regional-scale Wide Angle Camera bundles at roughly 100 meters per pixel.
- [7]
SomBench's pretraining corpus contains 1,000,113 Narrow Angle Camera bundles at roughly 1 meter per pixel.
- [8]
Each bundle can include surface imagery, terrain measurements and contextual data describing conditions such as illumination.
- [9]
The technical paper, posted on September 8th, evaluated the model across three task families: crater detection at regional and meter scales, segmentation of irregular mare patches and regression of lunar polar ice prospectivity.
- [10]
Across those tests the pretrained model matched or beat comparison systems initialized with ImageNet weights, along with an identical architecture that started from random weights.
- [11]
According to runtimewire, the missing pretraining code limits full reproducibility of the project.
- [12]
USRA detailed Slank's contribution on September 18th, describing the work of aligning observations collected by different instruments, at resolutions ranging from 1 meter to 20 kilometers per pixel, without discarding the physical context that gives each measurement scientific value.
- [13]
Slank served as a planetary-science subject-matter expert while working at NASA's Marshall Space Flight Center through USRA's Science and Technology Institute, moving between the science and modeling groups.
- [14]
For SomBench, the dataset and evaluation suite behind the model, Slank led development of a high-resolution crater benchmark based on Lunar Reconnaissance Orbiter Camera Narrow Angle Camera imagery, manually identifying over 49,000 craters and pairing the images with co-registered terrain models.
- [15]
Meter-scale crater detection is a task where small differences in lighting, slope or image resolution can resemble geological boundaries.
- [16]
The compute run required about 1,100 H100 GPU-hours, according to the model card.
- [17]
During her master's work at the University of Texas at El Paso, Slank used thermal-inertia data from the Lunar Reconnaissance Orbiter to investigate possible lunar cavities and lava tubes.
- [18]
The two pretraining pools sum to 1,963,722 bundles.
- [19]
Narrow Angle Camera bundles make up 50.9 percent of the pretraining corpus.
- [20]
About 1,100 H100 GPU-hours is 137.5 hours of wall clock on an eight-GPU node, or about 5.7 days.
- [21]
The pretraining run averages about two GPU-seconds per bundle.
- [22]
The resolutions being reconciled span a ratio of 20,000 to 1.
- [23]
The SomBench paper describes the dataset as containing more than 30 co-registered layers from ten instruments across four missions, while IBM's September 10th release says the unified dataset covers nine instruments across those four missions.
Sources
2 independent publishers whose own reporting we read for this story.
- runtimewire.comNASA and IBM released an open lunar AI model trained on two million data bundles
1 article · September 18, 2026
- the-decoder.comNASA and IBM's open source lunar model turns 17 years of orbiter data into a foundation for lunar science
1 article · October 4, 2026
Topics and entities
Topics
- Open-Weight Model Publishing
- Lunar Remote Sensing and Orbiter Imaging
- Foundation models for science
Entities
- IBM Research
- Lunar Reconnaissance Orbiter
- NASA
- NASA-IBM Lunar Foundation Model