IBM and NASA on September 10, 2026 announced the open-source release of the NASA-IBM Lunar Foundation Model, a multimodal, multi-resolution foundation model for lunar remote sensing, alongside SomBench, the co-registered lunar dataset built to train it.
In the announcement, dated from Yorktown Heights, New York, IBM described the system as one of the first publicly available foundation models for scientific exploration of the Moon, trained on an extensive lunar observation dataset curated by IBM and NASA researchers. Model weights are published on Hugging Face under the Apache-2.0 license, with fine-tuning code maintained in a NASA-IMPACT GitHub repository and downstream adaptation handled through TerraTorch. The model card frames the release as a reusable representation for lunar remote sensing: fine-tuning or LoRA-adapting the encoder for detection, segmentation, and dense regression on LROC imagery and co-registered terrain products, with modality subsets and patch grids that can differ from pretraining.
According to the announcement, researchers could apply the model to several lines of lunar study: predicting where ice may be present in permanently shadowed regions, which are difficult to observe yet may hold resources considered essential for a future Moon base; mapping Irregular Mare Patches, volcanic features scientists study to understand the Moon’s volcanic history and thermal evolution; and detecting, contextualizing, and classifying craters, work that reveals clues about terrain age and geology and helps NASA select safe landing sites while avoiding hazards such as steep slopes and boulders.
“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, adding that the model shows what is possible when AI is brought to NASA’s petabytes of scientific data.
IBM said the model extends an established collaboration between the two organizations and joins the Prithvi family of open foundation models available on Hugging Face, spanning geospatial, weather, heliophysics, and now the Moon. The announcement presents the open release as a broader vision in which researchers start from a shared model and adapt it to new tasks rather than building a new algorithmic system for every scientific question.
Juan Bernabe-Moreno, Director of IBM Research Europe, UK and Ireland, said the model gives scientists a foundation to explore the Moon at scale, connecting observations across instruments, revealing patterns that are difficult to see in isolation, and “providing an open platform the global research community can build on.”
Architecture and Pretraining
The model card describes a ViT-B encoder-decoder (768 dimensions, 12 layers, 12 attention heads) trained from scratch on SomBench: roughly 2 million co-registered lunar tile bundles spanning 11 modalities at two spatial scales, LROC Narrow Angle Camera imagery at about 1 meter per pixel and Wide Angle Camera imagery at about 100 meters per pixel. The system adapts the TerraMind masked-token recipe with two extensions. Per-tile acquisition geometry, including illumination angles, solar-frame anchors, and tile footprint, is tokenized as explicit encoder inputs; the authors state that lunar surface appearance is governed more by illumination geometry than by intrinsic surface variation. NAC- and WAC-anchored tiles also train together in a single mixed-batch loop at native resolution, so one set of weights covers both resolution families across a 100× scale gap.
Inputs span nine dense image-like layers plus two sequence-like context modalities: per-tile optical metadata with eight fields and static-map context with 28 fields drawn from Diviner, LOLA, Mini-RF, Kaguya, WAC, GRAIL, and Lunar Prospector products. A FlexiViT patch embedding allows the checkpoint to be fine-tuned at other patch sizes without retraining the backbone, and modality-wise tokenization allows modalities to be dropped or added at fine-tuning time. Pretraining ran on 16 H100 GPUs for 150,000 steps at a global batch of 1,536 in bf16, totaling roughly 1,100 GPU-hours, using nine modality-specific VQ-VAE tokenizers with FSQ quantization, a DDPM decoder, and a cross-entropy objective over discrete token vocabularies.
Reported Benchmark Results
IBM said the model exceeds widely used methods by up to 23% in identifying key geographic features on the Moon’s surface, including potential ice deposits, craters, and volcanic formations. Citing a NASA-IBM technical paper, the company reported that the model reduced error (RMSE) by up to 22% compared with the SwinV2-B (ImageNet) model when identifying areas with high potential for lunar ice; captured the extent of Irregular Mare Patches 3% better than SwinV2-B while using imperfect labels; and matched state-of-the-art models such as SwinV2-B on meter-scale crater detection while outperforming it by nearly 19% at roughly 100-meter context scale using half the training data.
The benchmark table on the model card reports polar ice prospectivity RMSE of 0.0293 with full fine-tuning, against 0.0377 for the best baseline (SwinV2-B) and 0.0397 for a random-initialization control, and identifies ice prospectivity as the widest margin. For Irregular Mare Patch segmentation it reports an IoU1 of 0.5709 against 0.5687 for ConvNeXtV2-B, and for Robbins crater detection on WAC imagery a mAP of 0.2581 using LoRA against 0.2420 for SwinV2-B; the card states that pretrained variants at 50% of the training data already match or exceed SwinV2-B trained on the full set. The authors caution that leaders on the NAC crater and Irregular Mare Patch benchmarks should be treated as comparable because margins are smaller than the spread across seeds. Benchmarks ran through TerraTorch with loaders, splits, augmentations, loss, and metrics held fixed across backbones, with baselines including ResNet-50, ViT-B MAE, ConvNeXt-B, ConvNeXt-V2-B, SwinV2-B, DaViT-B, DeepLabV3+, and SegFormer, reported as mean and standard deviation over five seeds.
SomBench Corpus and Stated Limitations
IBM said its scientists, working with NASA, built the first open-source lunar dataset of its kind, aggregating more than 30 spatially aligned layers from nine instruments across four missions, including NASA’s Lunar Reconnaissance Orbiter and GRAIL mission and the Japanese Aerospace Exploration Agency’s SELENE/Kaguya. The SomBench dataset card describes two tracks: WACLowRes, 963,609 tiles covering 51.2 kilometers at 100 meters per pixel drawn from 54,080 Wide Angle Camera Experiment Data Records and totaling 38 TB; and NACHighRes, 1,000,113 tiles covering 512 meters at 1 meter per pixel from 1,095 Narrow Angle Camera records, totaling 1.4 TB. The corpus is released under a CC BY 4.0 license, with the full dataset hosted on AWS. Splits are assigned at the Lunar Transverse Mercator zone level at a 75/15/10 target, tiles straddling two zones are dropped to prevent spatial leakage, and the test split is reserved for downstream evaluation.
The model card states the system is not a scientific-grade generative product, maintains no geodetic reference frame, and is not validated for operational decisions such as landing-site certification or hazard clearance; it also notes that ice-prospectivity outputs regress a knowledge-driven fuzzy-overlay map rather than measured ice. Listed limitations include ablation contributions not yet isolated and NAC pretraining restricted to 1,095 co-registered frames, while the dataset card notes heterogeneous spatial resolution and tens-of-meters absolute geolocation uncertainty. The work was supported by NASA under Award No. 80MSFC25M0084.
Credit: Source link
