NASA 与 IBM 开源月球科学基础模型，用 17 年轨道器数据预训练

The NASA-IBM Lunar Foundation Model makes decades of lunar observation data usable for machine learning. It's especially strong at predicting ice deposits at the poles and detecting craters.

"NASA has spent decades building an extraordinary scientific record of the Moon, but collecting data is only part of the job," says Kevin Murphy, NASA's chief science data officer. The data also has to be easier for scientists to use, he adds. NASA and IBM Research, working with several academic institutions, have now released the NASA-IBM Lunar Foundation Model. The two organizations describe it as one of the first open-source foundation models for lunar science.

Unlike task-specific algorithms, a foundation model is pretrained on large volumes of unlabeled data and can then be adapted to specific tasks with just a few labeled examples. The team sees this as the main benefit for lunar research, where observation data is plentiful but labels are scarce.

Nearly 2 million tile bundles from 17 years of observation

The team trained the model from scratch using SomBench, which they say is the largest co-registered multimodal lunar corpus to date. It contains nearly 2 million tile bundles across 11 modalities and two spatial scales. About 1 million high-resolution images come from the Narrow Angle Camera (NAC) at roughly 1 meter per pixel, while just under 964,000 multispectral images come from the Wide Angle Camera at 100 meters per pixel.

The bulk of the data comes from 17 years of observations by the Lunar Reconnaissance Orbiter (LRO). According to NASA, its data volume exceeds that of all other NASA planetary missions combined. Data from the GRAIL mission (gravity field), Lunar Prospector (hydrogen), and JAXA's Kaguya/SELENE probe (mineralogy) rounded out the collection.

In total, the dataset brings together more than 30 spatially aligned data layers from nine instruments and four missions. To prevent leakage between training, validation, and test data, the corpus is split geographically by map zones rather than randomly distributing individual tiles.

The model gets lighting conditions as input instead of guessing them

The model is based on TerraMind, a multimodal Earth observation model, but the researchers trained it from scratch instead of fine-tuning it. For each tile, the model receives the imaging geometry as explicit context, including illumination angles, sun position, and tile extent. On the Moon, lighting geometry shapes how the surface looks far more than the surface's actual properties do. Rather than forcing the model to correct for this from raw pixels, the team simply feeds it that information.

Diagram of the model architecture with four columns showing modalities like NAC and WAC optical data, elevation models, slope, aspect, UV, metadata, and static maps on the left, modality-wise tokenization in the middle, correlation learning with encoder and decoder, and downstream applications including multimodal generation, multimodal fine-tuning, and FlexiViT fine-tuning on the right. — The model breaks lunar images, elevation data, and imaging geometry into tokens and learns their relationships by predicting masked portions. | Image: NASA / IBM

The model also learns from high-resolution and coarse imagery together in a single training run, capturing both fine details and large-scale context. A technique called FlexiViT lets the same trained model adapt to tasks with different image patch sizes without retraining.

Ice prediction is where the model pulls ahead most

The team tested the model on crater detection at 100-meter and 1-meter scales and on predicting polar ice deposits. It also had to segment Irregular Mare Patches (IMPs), young volcanic features that challenge established timelines of lunar cooling. Across all tasks, the pretrained model matched or beat both common baselines and an architecturally identical control model with random initialization, according to the technical report.

Two polar maps of the Moon with four highlighted regions on the left, and three rows of color-coded prospectivity maps from blue to yellow on the right, comparing reference data with predictions from a ConvNeXt model and the NASA-IBM model for ice probability at the lunar poles. — The NASA-IBM foundation model closely reproduces the fine-grained patterns in reference maps of potential ice distribution at the lunar poles. | Image: NASA / IBM

The biggest gains showed up in ice deposit prediction. Permanently shadowed polar regions stay cold enough to preserve water ice for billions of years and are considered a potential resource for water, oxygen, and rocket fuel. The model cut prediction error by up to 22 percent compared to the best baseline, SwinV2-B, according to IBM.

Four lunar surface images with green boxes marking ground-truth craters and blue boxes marking detected craters. Top row shows two WAC images with a rille and few craters, bottom row shows two NAC images with densely distributed small craters. The NASA-IBM model is on the left, SwinV2-B on the right. — In these crater detection examples, the NASA-IBM model and the ImageNet-pretrained SwinV2-B perform similarly. All models do worse on high-resolution NAC images. | Image: NASA / IBM

For coarse-scale crater detection, the model beat SwinV2-B by nearly 19 percent, according to IBM. That number comes from training with only half the data, suggesting the model needs fewer labeled examples to perform well.

Part of the edge comes from how the model handles data

Even the randomly initialized control model beat five of seven baselines on ice prediction without any lunar pretraining. According to the researchers, part of the advantage comes from how the model handles different data types. It assigns each data layer its own processing path, while baseline models treat all inputs as stacked channels.

Grid of four NAC lunar surface images showing bright, irregularly shaped patches in the top row, with three rows below comparing segmentation results from Swin V2, ConvNeXt, and the NASA-IBM model, each showing green reference areas and purple prediction outlines. — When segmenting Irregular Mare Patches, the NASA-IBM model picks up areas in the second example that both baseline models miss. | Image: NASA / IBM

On meter-scale crater detection and IMP segmentation, the model roughly ties the strongest baselines, with differences falling within the variance between training runs. IBM claims a 3 percent lead over SwinV2-B on IMPs, but the results look more comparable than clearly better.

According to the researchers, LoRA, a lighter fine-tuning method that trains only a fraction of the parameters, kept pace with full fine-tuning across the board and did better on crater detection. Full fine-tuning only wins on the two smallest tasks.

Useful for analysis, but no substitute for physical measurements

Juan Bernabé-Moreno, director of IBM Research Europe, said the model connects observations across instruments and reveals patterns that are difficult to spot in isolation.

The model isn't suited for absolute geodetic positioning, the research shows. In generation tests, latitude and longitude were off by dozens of degrees in some cases, and elevation structures could appear with shifted absolute height values even when their shapes were reconstructed correctly.

The authors see it as a reusable foundation for downstream tasks, not a replacement for physical measurement instruments. Controlled experiments isolating the contribution of each innovation are still pending, and some test datasets are small.

Part of a growing model family

The model is publicly available on Hugging Face, the code is on GitHub, and it's integrated into the open-source toolkit TerraTorch. The team also released the ML-ready pretraining datasets and benchmark collections.

The model is part of the NASA-IBM "AI for Science" collaboration. Both organizations have been working on foundation models under a Space Act Agreement since early 2022. In August 2023, they released the first Prithvi model on Hugging Face, trained on Landsat and Sentinel-2 imagery of the entire contiguous US and adapted for flood and wildfire mapping.

IBM had announced Prithvi as part of its Watsonx platform. IBM developed TerraMind, the lunar model's base, in 2025 with ESA and Forschungszentrum Jülich for Earth observation.

Google Deepmind is pursuing a related approach with AlphaEarth Foundations, compressing optical satellite imagery, radar, lidar scans, and climate simulations into compact embeddings of Earth's surface.