Introducing IBM and NASA's new foundation model for the Moon

Details

IBM and NASA have released the NASA-IBM Lunar Foundation Model, an open-source model built on IBM's TerraMind architecture, originally co-developed with the European Space Agency for Earth observation. It is the first AI model to unify data across multiple sensors and spatial scales: from NASA's Lunar Reconnaissance Orbiter imaging at 1-meter-per-pixel resolution to GRAIL's gravity-field maps at 20 kilometers-per-pixel, plus data from Japanese lunar missions spanning decades. The core capability is cross-modal data fusion: the model learns correlations between measurement types and fills in missing or noisy sensor values. This follows the pattern of IBM's geospatial models Prithvi EO (Earth) and Surya (Sun), which are designed for fine-tuning on downstream tasks without full retraining from scratch. NASA has prioritized three initial applications: cataloguing smaller uncounted craters that pose navigation hazards; tracing ancient volcanic lava flows for geological research; and scanning permanently shadowed polar craters for water ice. Ice detection is especially mission-critical — water at the lunar poles could supply future astronaut crews with drinking water, oxygen, and rocket fuel. The release aligns with the Artemis program's goal of establishing a permanent human presence on the Moon. The model's direct operational deployment timeline for astronaut missions was not specified in the announcement.