Technology
NASA and IBM introduce open-source lunar foundation model for scientific exploration
Engadget reports that the organizations launched an open-source AI system to detect potential lunar ice and map craters.
The short version
- NASA and IBM published the open-source NASA-IBM Lunar Foundation Model on Hugging Face alongside a comprehensive lunar dataset.[Engadget]
- The model lowered errors by 23 percent relative to a baseline vision system when pinpointing potential lunar ice.[Engadget]
- Developers adapted training methods after conventional approaches struggled to differentiate between visually similar craters across missions.[Engadget]
Key facts
- NASA and IBM released the open-source NASA-IBM Lunar Foundation Model on Hugging Face.[Engadget]
- The model achieved a 23 percent reduction in errors compared to the SwinV2-B baseline when mapping areas of potential ice.[Engadget]
- In crater identification and classification, the system surpassed SwinV2-B by 19 percent using half as much training data.[Engadget]
- An accompanying open-source dataset compiles instrument data and imagery from NASA's LRO and GRAIL missions alongside Japan's SELENE mission.[Engadget]
What remains uncertain
- IBM Research official Juan Bernabé-Moreno noted that conventional training methods initially failed due to crater similarities, requiring an alternative wedge-based data separation approach.[Engadget]
Sources
Outlet counts describe coverage, not independent confirmation. Reports may share a wire service or original source.