NASA and IBM release open-source AI for lunar exploration

NASA and IBM have released a new artificial intelligence model designed specifically for lunar exploration. The system, known as the NASA-IBM Lunar Foundation Model, is available as open source, allowing researchers worldwide to download and use it. This release follows the recent success of the Artemis II mission, where astronauts completed a flyby of the Moon, reigniting interest in the agency’s plans for sustained lunar presence.

The primary goal of the model is to help scientists locate valuable resources, particularly water ice, on the lunar surface. By automating the analysis of high-resolution images, the tool aims to save researchers significant time in identifying promising areas for future landers. Engadget notes that this collaboration provides a crucial bridge between raw satellite data and actionable scientific insights.

Accuracy improves in key lunar tasks

In benchmark tests, the new model outperformed standard vision systems used in the industry. When tasked with identifying potential ice deposits, it reduced errors by 23 percent compared to the baseline tool, SwinV2-B. For crater classification, the model was 19 percent more accurate while requiring only half the amount of training data. This efficiency suggests that the system can process large volumes of lunar imagery with greater precision and lower computational costs.

The model’s capabilities were recently verified during an unexpected event. In August, a SpaceX Falcon 9 rocket booster crashed into the Moon. When IBM fed an image of the impact site to the model, it correctly identified the new crater despite it overlapping with an existing one. This success highlights the tool's ability to detect subtle changes in the lunar landscape, a critical skill for monitoring the Moon’s evolving surface.

Overcoming unique lunar imaging challenges

Training an AI for the Moon presents distinct difficulties not found in Earth observation. On Earth, the atmosphere scatters sunlight, filling shadows with ambient light that helps define shapes. On the Moon, shadows are pitch black and carry no visual information. This means a crater can look drastically different depending on the sun’s angle at the time of capture, making it hard for traditional AI models to recognize consistent features.

Standard training methods, which often involve hiding parts of an image to test reconstruction, failed initially because lunar craters look so similar to one another. IBM’s team described their first attempts as a disaster. To solve this, they divided the Moon into sections, like wedges of an orange, and strictly separated the training data from the testing data. This approach ensured the model learned consistent patterns without cheating by memorizing overlapping images.

Open data supports broader research

Beyond the model itself, the release includes a substantial open-source dataset. This collection contains tens of thousands of images and instrument data points, providing a foundation for other scientists to build their own tools. By making this data freely available, NASA and IBM aim to accelerate the pace of lunar research, allowing independent teams to develop new applications for exploring the Moon’s geology and resources.