NASA and IBM release open-source Lunar Foundation Model for moon science

The model, built on IBM’s TerraMind architecture and trained from scratch, ingests explicit lighting geometry for each tile and uses a FlexiViT technique to handle varying image patch sizes. In benchmark tests it cut polar‑ice prediction error by up to 22 % and improved coarse‑scale crater detection by nearly 19 % compared with the SwinV2‑B baseline, while requiring only half the labeled data. IBM and the researchers note that even a randomly initialized control model performed competitively on ice prediction, highlighting the value of the data handling approach.

All code, pretrained weights and the pretraining datasets are publicly available on Hugging Face and GitHub, and the model is integrated into the open‑source TerraTorch toolkit. Kevin Murphy, NASA’s chief science data officer, emphasized that the release makes decades of lunar observations easier for scientists to use, and Juan Bernabé‑Moreno of IBM Research Europe called the model a reusable foundation for downstream lunar research, not a substitute for physical measurements.