← Back to feed

NASA and IBM Research have released the NASA-IBM Lunar Foundation Model, an open-source AI trained from scratch on decades of multi-instrument lunar orbital data. The model is designed to help researchers analyse surface features, locate potential water ice deposits, and study volcanic formations — tasks that previously required domain-specific tools and manual data wrangling across incompatible instrument formats.

The Dataset: SomBench

The model was pretrained on SomBench, a multimodal lunar dataset assembled specifically for this project. SomBench contains nearly two million co-registered data bundles, where each bundle aligns observations from different instruments at the same lunar location.

• 11 measurement modalities — thermal, topographic, radar, multispectral, and gravity measurements among them
• 2 spatial scales to capture both local surface features and regional context
• Data from 4 lunar missions: the Lunar Reconnaissance Orbiter (LRO), GRAIL, Lunar Prospector, and SELENE/Kaguya

The co-registration problem — aligning observations from different instruments, different orbital geometries, and different time periods — was a significant engineering task. SomBench resolves it into a unified format that a single model can ingest.

Performance

The model preserves fine-scale patterns of lunar ice prospectivity that prior convolutional approaches smooth over. In a comparison against a ConvNeXt baseline at four locations near the lunar south pole, the NASA-IBM model reproduced the reference prospectivity map’s small-scale structure more accurately — the kind of detail relevant to identifying specific landing or excavation targets for future missions.

The model is available at arxiv.org/pdf/2609.13283 alongside the SomBench dataset paper (arxiv.org/pdf/2609.13277). Weights are on HuggingFace; inference runs through IBM’s TerraTorch geospatial AI framework on GitHub. The Universities Space Research Association (USRA) contributed the planetary science domain expertise and evaluation.

Why It Matters

Most planetary science AI is either single-instrument (a model trained on LRO data alone) or relies on manual feature engineering per dataset. A foundation model that can ingest 11 modalities simultaneously opens a path to automated cross-instrument analysis at scale — relevant not just for lunar science but as a template for AI on Mars, Europa, and other bodies where multi-instrument orbital data exists but synthesis is manual.

The open-source release is intentional: NASA and IBM are positioning SomBench and the foundation model as community infrastructure, the same way terrestrial geospatial AI has converged around shared datasets. Future mission teams can fine-tune the model on new instrument data without training from scratch.