Details
- IBM and NASA released the open-source NASA‑IBM Lunar Foundation Model in Sept 2026 to turn decades of multi-instrument lunar observations into a reusable AI foundation for Moon science.
- The collaboration involves IBM Research and NASA, extending their Prithvi family of open scientific foundation models into a new domain focused specifically on lunar exploration.
- The model improves identification of key lunar geographic features—ice deposits, craters, and volcanic formations—by up to 23% versus widely used baselines, while reducing error by up to 22% on lunar ice prediction and enhancing volcanic feature mapping efficiency.
- Alongside the model, IBM and NASA released a unified, machine-learning-ready lunar dataset aggregating over 30 spatially aligned layers from nine instruments across four missions, including LRO, GRAIL and JAXA’s SELENE/Kaguya.
- The Lunar Foundation Model joins the Prithvi geospatial, weather and heliophysics models as part of a broader strategy to use shared, open foundation models instead of bespoke algorithms for each scientific question, lowering fine-tuning costs and accelerating discovery.
Impact
The open-source Lunar Foundation Model and its unified dataset significantly lower the technical barrier for high-precision lunar mapping and resource prospecting, directly supporting Artemis-era plans for sustained human presence and infrastructure on the Moon. By extending the Prithvi-style foundation model approach beyond Earth observation to lunar science, IBM and NASA are reinforcing a trend toward domain-specific scientific foundation models that will likely shape AI-for-science R&D roadmaps and funding priorities over the next 12–24 months.
