NASA and IBM Release Open-Source Model for Studying the Moon

NASA and IBM Research, IBM's research division, released an open model for studying the Moon. It was trained on seventeen years of images from the Lunar Reconnaissance Orbiter, a NASA satellite that circles the Moon, plus data from three other lunar missions. It cut the error in predicting polar ice deposits by up to 22 percent compared with the previous best model, SwinV2-B. It also spots large craters almost 19 percent more accurately than that same model.

For a narrow task like lunar craters, a general-purpose model from Google or OpenAI will not help here. It simply has not seen enough similar images. A model built and trained specifically for this data beats the general-purpose model on its own turf, and it needs less training data to do it for the crater task. I see the same pattern in my own work. For a narrow, repetitive task, a model trained specifically for it usually beats a general-purpose model at that exact task.