NASA and IBM Launch Multi-Modal AI Foundation Model for Lunar Exploration

NASA and IBM have announced the open-source release of the NASA-IBM Lunar Foundation Model, the first artificial intelligence model designed to integrate lunar observations captured across different modalities, viewing angles, and spatial scales. Developed to support the Artemis program, the model harmonizes decades of data collected by US and Japanese missions, providing a robust computational tool for astronaut navigation, volcanic history investigation, and ice prospecting. The structural goal is to pave the way for establishing a long-term lunar base and preparing for future crewed missions to Mars.

The Challenge of Data Integration and Lunar Conditions

Planetary scientists face the historical challenge of handling massive volumes of multi-sensor data at drastically varying scales, such as the gravity field mapped by the GRAIL mission (20 kilometers per pixel) contrasting with the high-resolution imagery from the Lunar Reconnaissance Orbiter (1 meter per pixel). Beyond the disparity in resolutions, the absence of a thick atmosphere and the Moon’s lighting cycle—which alternates between two weeks of continuous sunlight and two of darkness—create long shadows and severe visual distortions. These conditions complicate the interpretation of thermal data, where temperatures violently swing from 121°C in direct sunlight to -246°C in shadowed craters. Until now, traditional simulations to increase the resolution of these thermal and visual maps required massive time and computational resources, limiting the speed of analysis for selecting landing sites.

TerraMind-Based Architecture and LoRA Optimization

To unify these divergent measurements, software engineering teams from IBM and NASA adapted a version of TerraMind, an Earth-observation model originally developed in partnership with the European Space Agency (ESA). The system is capable of learning cross-modal correlations to fill in missing or noisy sensor values. Under a fine-tuning approach, researchers applied lightweight low-rank adapters (LoRAs), a technique that allowed training the system while keeping 90% of the base model’s weights frozen. Despite this computationally low-cost approach, the performance surpassed traditional architectures: in identifying polar craters likely to contain ice, the model reduced the error rate by 22% compared to a state-of-the-art SwinV2 transformer. In crater detection at coarser resolutions (100 meters per pixel), accuracy was nearly 19% higher than the equivalent model, using only half the training data.

“The next lunar breakthrough could come not from a single new instrument, but from algorithms that allow many instruments — and eventually, computing paradigms — to work together.”

Scientific Impact and the Future of Hybrid Computing

The model’s ability to quickly isolate promising sites will optimize the prospecting of the estimated 600 million metric tons of frozen water at the lunar poles, essential resources for the local production of drinking water, oxygen, and fuel. In the field of planetary geology, the AI achieved a 3% higher accuracy in mapping the extents of irregular mare patches (IMPs), volcanic formations whose exact age remains a subject of scientific debate. The release of this model also signals a transition in high-performance scientific computing workflows. In the long term, artificial intelligence integration will organize observations and generate rapid approximations, while classical supercomputers process core workloads, paving the way for quantum processors to handle the most difficult simulation and optimization problems.