NASA and IBM presented NASA-IBM Lunar Foundation Modelan open AI system for analyzing the lunar surface. The model is available for download at Hugging Face along with a dataset on which other researchers can train their own systems. The release comes as the agency prepares for the next missions of the Artemis program.
The reason for the work was the growing interest in the Moon after the flight of Artemis II, the first manned mission to the satellite since 1972. Now NASA wants to give scientists tools that will speed up preparations for the landing.
According to Juan Bernabe-Moreno, director of IBM Research Europe in the UK and Ireland, the fundamental model is capable of solving several types of problems at once. First, NASA and IBM discovered that the system was particularly good at identifying areas of the lunar surface where ice might lie.
The developers compared the model’s predictions with a map built using a published scientific algorithm that took into account terrain, thermal and other environmental data. The system was then tested against SwinV2-B, a Microsoft-trained computer vision model for high-resolution image processing. NASA and IBM’s model reduced errors by 23 percent.
SwinV2-B chosen as a starting point, since this system is often the basic solution for image analysis problems. In a separate test for crater detection and classification, the new model outperformed its competitor by 19 percent, while using half the training data.
It was possible to test the capabilities of the system in real conditions after the SpaceX Falcon 9 rocket fell on the Moon on August 5. When IBM specialists loaded the model with a snapshot of the impact site, it correctly recognized a fresh crater, despite the fact that it almost completely overlapped with an existing one.
Juan Bernabe-Moreno, director of IBM Research Europe in the UK and Ireland:
It worked fantastic
Training the model proved to be a challenging task. Earth’s atmosphere scatters sunlight and fills shadows with diffuse illumination, blurring their boundaries, so satellite images of the planet look clean and informative. On the Moon, shadows are completely black with sharp edges, and the shadowed pixels do not carry any information at all.
On the Moon you have the Sun at different times of the day, it illuminates the surface and cuts out shadows, and these shadows need to be dealt with somehow
As a result, the same crater can look completely different in two images. Additional difficulties were created by the standard approach to training computer vision models, in which part of the image is removed, and the system is asked to restore the missing fragment using the learned patterns.
You can remove 90 percent of the crater and use the remaining 10 percent to ask the model to restore the remaining 90 percent.
This method is not suitable for the Moon, since many craters look almost the same when photographed from orbit. Bernabe-Moreno called the first attempts to train the system in the traditional way a complete disaster. Researchers found a non-standard solution and divided the surface of the satellite into sectors.
We took the Moon and divided it into wedges, like an orange, and completely separated the training wedges from the test wedges. This way we were able to provide greater consistency to the model.
A special value of the release is the first of its kind open data set for lunar research. It included tens of thousands of images and instrument readings from NASA’s Lunar Reconnaissance Orbiter and GRAIL missions, as well as the Japanese SELENE spacecraft.
One of the reasons we’ve never seen a comprehensive, powerful lunar model before is because we didn’t have the data organized in the right way.
Scientists have found that an earthly fungus can live up to 7 days near the south pole of the Moon.
NASA risks losing unique data about the edge of the solar system due to savings of $2 million
Organizing in this case meant creating a grid, where each pixel in the image corresponds to other types of data that the model can take into account when processing. According to Bernabe-Moreno, this was the main contribution of the work to science.
This in itself is a huge scientific contribution because, model aside, the community now has a coordinated dataset of over two million points. And we know that AI models are being replaced by other models. Data is what drives the AI industry
There is no exact date for the Artemis III mission yet.
