Ay araştırmaları için açık bir yapay zekâ modeli

NASA and IBM have released an open-source foundation model designed to help scientists examine the Moon’s surface. Instead of being trained for only one narrowly defined task, the system learned broad patterns from roughly two million image tiles, mainly collected by the Lunar Reconnaissance Orbiter. Researchers can then fine-tune it with relatively small labelled datasets for specific purposes. Tests show that the model can map craters, identify unusual volcanic formations, detect changes between observations and estimate where ice may remain stable near the lunar poles. It matched or surpassed several strong comparison models, with its clearest advantage appearing in predictions about polar ice. The release also includes training datasets and benchmarks, allowing researchers to reproduce results and improve the technology. NASA nevertheless notes a practical limitation: changing illumination between spacecraft passes can make small craters harder to compare. The model is therefore a flexible scientific aid, not an automatic substitute for expert interpretation.