NASA and IBM Release Open-Source Lunar Foundation Model to Map Ice and Craters

IT Home reported on September 10 that this spring, NASA's Artemis II mission completed a historic lunar flyby, marking humanity's first crewed flight around the Moon since 1972. On April 6, astronauts Reid Wiseman, Christina Koch, Victor Glover, and Jeremy Hansen traveled farther into deep space than anyone before them. Now, as NASA prepares for subsequent Artemis missions, it is collaborating with IBM to build brand-new lunar research tools for scientists.

On Thursday local time, the two organizations jointly released the NASA-IBM Lunar Foundation Model. This is an open-source artificial intelligence system available for download on Hugging Face. Dr. Juan Bernabé-Moreno, Director of IBM Research for Europe, the UK, and Ireland, explained that as a foundation model, it can perform a variety of tasks. First and foremost, NASA and IBM found that the model is particularly adept at identifying areas on the lunar surface that may contain ice.

IT Home noted that both parties conducted a benchmark test: comparing the model's predictions with a lunar map generated through established scientific workflows that integrate topography, temperature, and other environmental data. The research team also benchmarked the model against SwinV2-B, a high-resolution image vision processing model developed by Microsoft. Ultimately, the NASA-IBM model reduced identification error by 23%. SwinV2-B was selected as the reference because it is widely used as a baseline model for various image analysis tasks.

NASA and IBM also evaluated the model's ability to identify and classify craters. In this test, using only half the training data, the new model outperformed SwinV2-B by 19%. On August 5, a SpaceX Falcon 9 rocket stage impacted the Moon, offering the research team an opportunity to validate the model's real-world capability. When IBM fed the post-impact imagery into the model, it accurately identified the newly formed crater, even though it heavily overlapped with an existing one. Bernabé-Moreno commented: "The result was outstanding; the model identified it successfully on the first try."

Bernabé-Moreno noted that training the model itself was full of challenges. He explained: "In Earth remote sensing, we get pristine imagery. Earth's atmosphere scatters sunlight, and ambient light fills in shadowed areas, softening their edges. But shadows on the Moon are razor-sharp and pitch-black. When photographed, pixels in shadow regions carry no usable information." He added: "On the Moon, the sun's angle varies across different times, alternately illuminating the surface and casting stark shadows. The model must handle these dramatic lighting shifts. The exact same crater can look completely different in different photos."

Another hurdle came from the conventional training and testing methodology used for computer vision models. The standard practice typically involves masking portions of an image and having the model reconstruct the missing parts based on learned visual patterns. Bernabé-Moreno gave an example: "We could mask 90% of a crater image, leave just 10%, and have the model reconstruct the remaining 90%." This approach failed in the lunar context. The reason is simple: in orbital imagery, vast numbers of craters look remarkably alike.

Bernabé-Moreno said that IBM initially attempted to train the model using conventional methods, which resulted in "complete failure." The research team ultimately devised an innovative solution: "We sliced the Moon like an orange into multiple segments, completely separating the training set segments from the test set segments. This allowed model training to maintain solid consistency."

A major reason NASA and IBM have high expectations for this release is the accompanying launch of the world's first open-source dataset of its kind, which other researchers can use to develop new models. The dataset brings together tens of thousands of images and instrument readings sourced from NASA's Lunar Reconnaissance Orbiter (LRO), the Gravity Recovery and Interior Laboratory (GRAIL) mission, and Japan's SELENE (Kaguya) lunar orbiter.

Bernabé-Moreno said: "A major reason why a comprehensive, powerful lunar foundation model had never existed before was the lack of properly curated accompanying data." Here, data curation refers to building a grid system where every pixel in an image aligns with other dimensions of sensor data accessible to the model. "This accomplishment alone is a major scientific breakthrough. Setting the model aside, the scientific community now possesses a co-registered, aligned dataset containing over 2 million data points." He added: "AI models will always be superseded by newer ones, but data is the true foundation sustaining the continued evolution of AI."