Following the historic Artemis II flyby this spring, NASA and IBM have announced the release of the NASA-IBM Lunar Foundation Model, an open-source artificial intelligence system designed to aid in lunar research. The model is now available for download on Hugging Face, accompanied by a novel dataset that researchers can use to develop their own applications.
Dr. Juan Bernabé-Moreno, director of IBM Research Europe, UK, and Ireland, highlighted the model’s proficiency in detecting subsurface ice. In benchmark tests against a scientific workflow incorporating terrain and thermal data, the new model reduced errors by 23 percent compared to SwinV2-B, a Microsoft-trained vision system commonly used as a baseline for image analysis. Additionally, when tasked with identifying and classifying craters, the NASA-IBM model outperformed SwinV2-B by 19 percent while requiring half the training data.
The system’s capabilities were recently validated during the August 5 impact of a SpaceX Falcon 9 rocket stage. When IBM fed an image of the crash site to the model, it successfully identified the new crater despite it overlapping with an existing formation, achieving accurate detection on its first attempt.
Training the model presented unique challenges due to the Moon’s harsh lighting conditions. Unlike Earth, where the atmosphere scatters sunlight and softens shadows, lunar shadows are pitch black and sharply defined, causing craters to appear drastically different depending on the time of day they are photographed. Traditional computer vision training methods, which involve masking parts of an image and asking the model to reconstruct the missing data, proved ineffective because many craters look similar from orbit. Bernabé-Moreno described these initial attempts as a “complete disaster.” The team ultimately solved this by dividing the lunar surface into wedge-shaped sections, ensuring that training and testing data were completely separated to maintain consistency.
A key component of the release is the accompanying open-source dataset, which aggregates tens of thousands of images and instrument data from NASA’s Lunar Reconnaissance Orbiter (LRO) and Gravity Recovery and Interior Laboratory (GRAIL) missions, as well as Japan’s Selenological and Engineering Explorer (SELENE). Bernabé-Moreno emphasized that the dataset organizes this information into a grid where every pixel of imaging data corresponds with other modalities, creating a co-registered resource with over two million data points.
“One of the reasons we’ve never seen a comprehensive, powerful lunar model before is that we didn’t have the data organized in the right way,” Bernabé-Moreno stated, noting that while AI models may be superseded over time, the underlying data remains a critical asset for the industry.
I hope this open-source approach means we see lots of new tools built on top of it, not just one big corporate product.
Finally, someone organized the LRO and GRAIL data into a single grid. That dataset alone might be worth more than the model.
The part about training being a ‘complete disaster’ is hilarious. At least they figured out that wedge-shaped segmentation trick in the end.
Reduced errors by twenty-three percent is impressive. I wonder how it handles the polar regions with their perpetual shadows?
This is a game-changer for finding water ice. Accurate subsurface detection will make future Artemis missions so much safer.