For 17 years NASA’s Lunar Reconnaissance Orbiter has photographed the Moon at roughly one meter per pixel, producing more data than every other NASA planetary mission combined. On September 10, the agency and IBM Research released the model that read it — along with the weights, the fine-tuning code, and the nearly 40-terabyte training corpus underneath it.
The NASA-IBM Lunar Foundation Model is among the first open-source foundation models built specifically for lunar science. The weights sit on Hugging Face under an Apache 2.0 license, fine-tuning code lives in a NASA-IMPACT GitHub repository, and downstream adaptation runs through TerraTorch, IBM’s open Earth-observation toolkit. The pre-training corpus, SomBench, ships under CC BY 4.0 on AWS.
“NASA has spent decades building an extraordinary scientific record of the Moon, but collecting data is only part of the job,” said Kevin Murphy, chief science data officer and acting chief data and AI officer at NASA Headquarters. “We also have to make data easier for scientists to explore and use.”
What is actually in the box
The architecture is modest: a ViT-B encoder-decoder, 768 dimensions, 12 layers, 12 attention heads, trained from scratch. The engineering is in the data. SomBench holds roughly 2 million co-registered tile bundles spanning 11 modalities at two scales — 1,000,113 tiles at about 1 meter per pixel from the Narrow Angle Camera, and 963,609 multispectral tiles at about 100 meters per pixel from the Wide Angle Camera. The high-resolution track totals 1.4 TB, the coarse track 38 TB. Layered on top are more than 30 spatially aligned products from nine instruments across four missions: Diviner, LOLA and Mini-RF on LRO, plus GRAIL gravity, Lunar Prospector neutron data and JAXA SELENE/Kaguya imagery.
Two departures from the standard recipe matter. The team adapted the TerraMind masked-token approach but tokenized per-tile acquisition geometry — illumination angles, solar-frame anchors, tile footprint — as explicit encoder inputs, reasoning that lunar appearance is governed more by illumination than by intrinsic surface variation. Second, narrow-angle and wide-angle tiles train together in one mixed-batch loop at native resolution, so a single set of weights spans a 100-fold scale gap. Pre-training ran on 16 H100 GPUs for 150,000 steps at a global batch of 1,536 — about 1,100 GPU-hours, a rounding error next to any frontier model run.
The advertised result is error cut by up to 23 percent. The underlying numbers are narrower. Polar ice prospectivity is the standout: root-mean-square error of 0.0293 with full fine-tuning against 0.0377 for the best baseline, SwinV2-B, roughly 22 percent better. Crater detection on wide-angle imagery hit 0.2581 mAP using LoRA against 0.2420 for SwinV2-B, about 19 percent better on half the training data. Segmentation of irregular mare patches — young-looking volcanic features that complicate established timelines for lunar cooling — improved only marginally, 0.5709 IoU against 0.5687. The authors note that on the narrow-angle crater and mare-patch tasks, margins fall inside the spread across random seeds.
“Uncovering the mysteries of the Moon requires an ability to learn from an extraordinary volume of scientific data,” said Juan Bernabe-Moreno, director of IBM Research Europe, UK and Ireland, describing the release as “an open platform the global research community can build on.”
The Artemis connection is obvious, and the model card goes out of its way to contain it. Crater maps and slope hazards bear on landing-site selection for a program targeting a crewed return in 2028, and permanently shadowed regions near the south pole are why ice prospecting matters at all. But the documentation states the system keeps no geodetic reference frame and is not validated for landing-site certification or hazard clearance. Ice outputs regress a knowledge-driven fuzzy-overlay map, not measured ice. Narrow-angle pre-training drew on just 1,095 source frames — the asterisk on the two-million figure: these are tiles, not distinct photographs.
Analysis: the small, boring, shared model
Strip away the Moon and this release is an argument about how scientific AI should be built. Eleven hundred GPU-hours and an 86-million-parameter backbone would be unremarkable at any AI lab. What is scarce here is co-registration — the unglamorous work of aligning gravity, topography, radar, thermal and multispectral products from four missions across three decades onto a common grid, then splitting it by Lunar Transverse Mercator zone so spatial leakage does not inflate scores. Domain-specific science models do not win on scale. They win on whether somebody did that work once and published it.
Which is why open weights matter more in planetary science than in most fields. The binding constraint on lunar machine learning is labels: there is exactly one Moon, a handful of mapped mare patches, and no prospect of crowdsourcing ground truth. An encoder that reaches baseline parity on half the labeled data turns a three-year annotation project into a semester. The shared benchmark may outlast the model — before SomBench, every lunar detection paper reported on its own splits, making the literature essentially incomparable. Holding loaders, augmentations, losses and metrics fixed across eight baselines over five seeds is the more durable contribution.
It also fits a pattern. This is the fourth act in the NASA-IBM collaboration, after the Prithvi geospatial family and the Surya heliophysics model in 2025: NASA supplies the archive, IBM supplies the training recipe, Hugging Face supplies distribution. That division of labor is now the template for agency-scale scientific AI, and the permissive license means a graduate student in Bengaluru or Munich can fine-tune it on Chandrayaan data next week without a data-sharing agreement.
What to watch next: whether the technical report survives peer review intact; whether outside groups adapt the encoder to non-NASA lunar datasets, the real test of a general representation; whether narrow-angle pre-training expands beyond 1,095 frames; and whether any Artemis or CLPS site-selection workflow adopts a model its own card currently disclaims for that use. Which body gets one next? The same recipe points straight at Mars, where the orbital archive is larger still.
“NASA has spent decades building an extraordinary scientific record of the Moon, but collecting data is only part of the job. We also have to make data easier for scientists to explore and use.”— Kevin Murphy, Chief science data officer, NASA Headquarters