NASA and IBM released an open lunar AI model trained on two million data bundles

Why it matters

The project shows how domain experts can turn fragmented public science data into reusable AI infrastructure, while the missing pretraining code limits full reproducibility.

Planetary scientist 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.

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. USRA detailed Slank's contribution on September 18th, describing the less marketable work beneath the foundation-model label: 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.

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. She moved between the science and modeling groups, connecting lunar research questions with decisions about the model's inputs, applications and evaluation.

The model's central challenge was reconciling heterogeneous Moon data. NASA and its partners have accumulated observations from instruments with different resolutions, coordinate systems, coverage and measurement conventions. A model can ingest those files while still learning relationships that make little physical sense. Slank's job was to help keep the machine-learning pipeline attached to the Moon it was supposed to represent.

A planetary scientist inside the training loop

Slank's background tracks closely with the model's intended uses. She earned a Ph.D. in space and planetary science from the University of Arkansas, where she studied Martian salts, liquid water and habitability. During her master's work at the University of Texas at El Paso, she used thermal-inertia data from the Lunar Reconnaissance Orbiter to investigate possible lunar cavities and lava tubes. She later held a postdoctoral fellowship at the Lunar and Planetary Institute.

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. She manually identified over 49,000 craters and paired the images with co-registered terrain models.

The annotation work supplied a ground truth for testing meter-scale crater detection, a task where small differences in lighting, slope or image resolution can resemble geological boundaries. It also illustrates the labor hidden by the phrase "pretrained from scratch." The compute run required about 1,100 H100 GPU-hours, according to the model card, but the useful evaluation depended on a scientist drawing tens of thousands of crater boundaries first.

The SomBench paper describes the dataset as containing more than 30 co-registered layers from ten instruments across four missions. IBM's September 10th release instead says the unified dataset covers nine instruments across those four missions. SomBench's pretraining corpus contains 963,609 regional-scale Wide Angle Camera bundles at roughly 100 meters per pixel and 1,000,113 Narrow Angle Camera bundles at roughly 1 meter per pixel. Each bundle can include surface imagery, terrain measurements and contextual data describing conditions such as illumination.

Illumination is especially important on the Moon, where shadows can dominate an image even when the underlying terrain changes little. 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. It also trains across both image scales using one set of weights, covering a 100-fold resolution gap.

Promising benchmarks, narrow claims

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.

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. IBM reported up to a 22% reduction in error for ice-prospectivity estimates against a SwinV2-B baseline. At the regional crater scale, IBM said the model produced nearly 19% better performance while using half as much training data. Meter-scale crater results were comparable to the strongest baseline.

Those figures come from the project team's own arXiv preprint and have not undergone peer-reviewed validation. They show performance on defined benchmarks, rather than use during an active lunar mission or certification for landing-site decisions.

The distinction is clearest around ice. The model estimates ice prospectivity using a knowledge-driven reference map. It does not detect or confirm accessible deposits. The model card also warns that generated fields are not calibrated scientific predictions, that location estimates can drift by tens of degrees and that the system has not been validated for landing-site certification or hazard clearance.

The practical near-term use is narrower: researchers can start with a representation already trained on lunar data, then adapt it to a particular detection, segmentation or regression problem with fewer labels. For scientific fields where expert annotation is expensive, that label efficiency may prove more consequential than a single benchmark lead.

Open weights, with a reproducibility gap

NASA and IBM released the model weights through Hugging Face and published fine-tuning and inference code on GitHub under an Apache-2.0 license. The repository includes configurations for crater detection, irregular mare patch segmentation and ice-prospectivity regression.

The GitHub repository explicitly says the pretraining code is not included. Researchers can inspect and adapt the checkpoint, reproduce downstream experiments and build applications around it, but they cannot recreate the full pretraining run from the public repository alone. That leaves the release open for use while limiting end-to-end reproducibility of the model-building process.

NASA and IBM have positioned the project within a broader effort to replace one-off scientific algorithms with reusable models trained across large public datasets. The lunar model extends work that IBM and NASA have already done on geospatial, weather and heliophysics models. Here, the institutional bet is that observations collected for separate missions can become shared infrastructure for questions that researchers have yet to formulate.

Slank's contribution shows what that infrastructure requires. The model began with data engineering, geological judgment and painstaking labels, then added the GPUs. Its value will depend on whether other planetary scientists can use those foundations to produce findings that survive beyond the benchmark table.