For 17 years, NASA's Lunar Reconnaissance Orbiter has been circling the Moon, amassing more observational data than all other NASA planetary missions combined. But abundant data is not usable data: planetary scientists either pored over maps and imagery by hand or trained low-resolution, task-specific models from scratch for each question — computationally expensive and often short on scientific accuracy. On September 10, NASA and IBM announced the open-source release of the Lunar Foundation Model, changing that path.
Polar ice prospectivity: error down up to 22% versus baseline
Of the three downstream evaluations in the official technical report, the headline is polar ice prospectivity: in permanently shadowed regions — the hardest environments to observe — the model reduces root-mean-square error in identifying high-potential ice areas by up to 22% compared with the widely used SwinV2-B (ImageNet) baseline. The model card lists the raw numbers, 0.0293 versus 0.0377, matching the press release. Lunar ice means water and oxygen — feedstock for a future Moon base and fuel for Mars missions — so this capability matters directly.
On crater detection, the press release says that at context scale (about 100 meters) the model leads SwinV2-B by nearly 19% using only half the training data; at meter scale it is roughly on par with the strongest baseline — 0.1543 versus 0.1552 in the model card, marginally behind. For irregular mare patches, the release claims an edge of about 3%, but the model card itself pours cold water: the margin is smaller than seed-level spread, so treat it as a tie. The NASA page also shows a case where a SpaceX rocket body punched a fresh crater onto the surface; on post-impact imagery that was excluded from pretraining, the model flagged the newly formed crater — exactly the change-detection capability that future lunar-activity monitoring needs.
One ViT-B — why it works
The architecture is not radical: a ViT-B encoder-decoder (768-dim, 12 layers, 12 heads) following the masked-token recipe of TerraMind, an Earth-observation foundation model. The real additions are two lunar-specific moves. First, per-tile illumination geometry — solar incidence and azimuth angles, tile footprints — is explicitly sequence-tokenized as context, because lunar surface appearance is governed more by illumination geometry than by intrinsic surface variation; the dominant confound is handed to the model instead of being left for it to recover. Second, tiles at 1-meter (NAC) and 100-meter (WAC) resolution train in a single mixed batch at native resolution, so one set of weights spans both resolution families across a 100x scale gap.
The training corpus, SomBench, aggregates more than 30 spatially aligned layers from nine instruments across four missions into roughly 2 million co-registered tile bundles — over 1 million 1-meter Narrow Angle Camera images and about 964,000 100-meter multispectral images, plus data from NASA's GRAIL gravity mission, Lunar Prospector, and JAXA's SELENE. Training cost: 16 H100 GPUs, about 1,100 GPU-hours. This is not an arms race in compute; it is a data-engineering exercise, and aligning multi-mission, multi-instrument data into machine-learnable form is the genuinely expensive part.
The cold water in the model card
The open-source trio is complete: weights, fine-tuning code integrated with the TerraTorch toolkit, and the SomBench dataset and benchmark collections, all under Apache-2.0. But the Out-of-scope section is blunt. The model maintains no geodetic reference frame — generated lat/lon can drift by tens of degrees. Ice-prospectivity outputs regress a knowledge-driven fuzzy-overlay map, not measured ice. Meter-scale pretraining is limited to 1,095 frames with co-registered 3-meter stereo DTMs — globally distributed, but not globally dense. And it is not validated for operational decisions such as landing-site certification. In short: a reusable representation backbone, not a finished scientific instrument.
So what
Kevin Murphy, NASA's chief science data officer, framed it this way: collecting data is only part of the job — the real opportunity is turning large-scale data into new discoveries. The family this model joins, Prithvi for Earth observation and Surya for heliophysics, and now the Moon, represents an emerging playbook: instead of building a new algorithm for every scientific question, maintain a shared backbone and fine-tune it to new tasks with small labeled sets. For AI-for-Science practitioners the lesson is direct: in domain foundation models, the decisive factor is often not parameter count or compute, but the patience to curate domain data into learnable form. A bit over a week after the announcement, monthly downloads sit at just above two thousand — whoever fine-tunes it into something unexpected may simply not have shown up yet.