NASA and IBM Release Open-Source Lunar AI Model and Dataset

The model was trained on a large multimodal NASA dataset, which is being released alongside it. TechRadar reported that the release is meant to give researchers wider access to advanced AI systems and support lunar exploration. The dataset is the first unified, publicly available cache ready for machine learning, according to TechRadar, bringing together more than 30 spatially aligned layers from nine instruments across four missions, including tens of thousands of images and maps from NASA's Lunar Reconnaissance Orbiter and GRAIL mission.

Engadget reported that the dataset also incorporates instrument data from Japan's Selenological and Engineering Explorer, or SELENE. It quoted Dr. Juan Bernabé-Moreno, director of IBM Research Europe, UK and Ireland, saying that one reason a comprehensive lunar model had not appeared before was that the data had not been organized in the right way. He said the community now has a co-registered dataset with more than two million data points.

The model is intended to make it easier to study petabytes of lunar surface data for hazards and resources such as ice deposits and craters. TechRadar reported that scientists currently rely on manual analysis or low-resolution, task-specific AI models, which can be computationally demanding and lack accuracy for detailed geographic analysis. A foundation model lets them adapt a single system to investigate a range of lunar geologic features instead of building a new model for every problem.

Tests described by Engadget showed the model was particularly adept at identifying possible ice deposits. NASA and IBM compared its predictions against a map made using a published scientific workflow that incorporates terrain, thermal and other environmental data, then tested it against SwinV2-B, a Microsoft-trained vision system. Engadget reported that the NASA-IBM model reduced errors by 23 percent.

The model also outperformed SwinV2-B in identifying and classifying craters by 19 percent while using half the training data, according to Engadget. Engadget reported that when a SpaceX Falcon 9 rocket crashed into the Moon on Aug. 5, IBM fed an image of the impact to the model, which correctly identified the crash site as a new crater despite its close overlap with an existing crater. Bernabé-Moreno said it "worked fantastically" and identified the site on its first attempt.

Training the model presented challenges. Bernabé-Moreno told Engadget that Earth observation images are relatively pristine because the atmosphere scatters sunlight and softens shadows, while lunar shadows are knife-edged and pitch black, meaning shadowed pixels carry no information. A crater can look different from one image to the next. The team's traditional training approach was a "complete disaster," he said, until researchers divided the Moon into wedges like an orange and separated training wedges from testing wedges.

IBM has worked with NASA for more than five decades, including on the Apollo missions, TechRadar reported. The company said the model could help future astronauts navigate safely and find essential resources. In a statement to TechRadar, Bernabé-Moreno said: "Uncovering the mysteries of the Moon requires an ability to learn from an extraordinary volume of scientific data. The NASA-IBM Lunar Foundation Model gives scientists a foundation to explore the Moon at scale, connecting observations across instruments, revealing patterns that are difficult to see in isolation, and providing an open platform the global research community can build on."

The work is focused on several lunar features. Ice deposits are considered vital because water and oxygen would be needed for a human base and for producing rocket fuel for future Mars missions, according to TechRadar. Volcanic features known as irregular mare patches can help scientists understand the Moon's volcanic history and thermal evolution, as well as identify potential landing and surface operation sites. Studying craters can reveal information about the age of terrains, their geology and the chemical composition of the early lunar interior, while also helping identify safe landing sites without steep slopes and boulders.

Engadget reported that the release follows NASA's Artemis II mission, the first crewed flight to the Moon since 1972, which completed a lunar flyby in April. Astronauts Reid Wiseman, Christina Koch, Victor Glover and Jeremy Hansen flew farther from Earth than any humans before them, according to Engadget. As NASA prepares for the next Artemis mission, the agency and IBM are offering the model and dataset as open tools for scientists.