Most AI foundation models are trained on language or general imagery. NASA↗ and IBM↗ have taken the same idea into planetary science with a model trained specifically to understand the Moon.
The Lunar Foundation Model combines more than 30 layers of data from nine instruments flown on four NASA missions. The model can learn relationships across imaging, topography and other measurements and then be adapted to tasks such as identifying craters or possible ice deposits.
In benchmark testing, NASA and IBM report improvements of as much as 23% over existing methods. The practical use is straightforward: future lunar missions need better maps of safe landing terrain and resources, particularly water ice that could support crews and potentially be processed into oxygen or rocket propellant.
NASA and IBM released an open-source lunar foundation model trained on more than 30 data layers from nine instruments across four missions. In benchmarks, the system improved identification of lunar features by as much as 23%.