AI news ·
NASA and IBM release open-source lunar foundation model trained on 17 years of orbiter data
NASA and IBM open-sourced a lunar AI model that cuts polar-ice prediction error by up to 22% and needs only half the labeled data.
NASA and IBM Research have open-sourced a new AI model trained on nearly 2 million tile bundles from 17 years of lunar orbiter missions. The NASA-IBM Lunar Foundation Model reduces polar-ice prediction error by up to 22% and improves coarse-scale crater detection by about 19% compared with a SwinV2-B baseline, while requiring only half the labeled data to achieve those results.
The model, built on IBM's TerraMind architecture and trained from scratch, processes explicit lighting geometry for each tile and uses a FlexiViT technique to handle varying image patch sizes. All code, pretrained weights, and pretraining datasets are publicly available on Hugging Face and GitHub, and the model is integrated into the open-source TerraTorch toolkit.
What the model does differently
The TerraMind architecture ingests multimodal data - visual imagery paired with lighting geometry metadata - rather than treating lunar images as standalone photographs. This approach proved valuable even without trained weights: a randomly initialized control model performed competitively on ice prediction, which the researchers said underscores the importance of the data handling method itself.
For crater detection, the model delivered a roughly 19% improvement over the SwinV2-B baseline at coarse scales. On polar-ice prediction, error dropped by up to 22%. Both benchmarks used half the labeled data that the baseline required, which matters in planetary science where annotated training data is scarce and expensive to produce.
Designed for reuse, not replacement
Kevin Murphy, NASA's chief science data officer, said the release makes decades of lunar observations easier for scientists to use. Juan Bernabé-Moreno of IBM Research Europe called the model "a reusable foundation for downstream lunar research, not a substitute for physical measurements." The distinction matters: the model accelerates analysis but does not replace the instruments that collect raw data.
The model page and full paper are available through the NASA-IBM Lunar Foundation Model page. Researchers can also access the model through TerraTorch, the open-source toolkit where it is now integrated alongside other geospatial AI tools.
Why this matters for science and research professionals
Planetary scientists and geospatial researchers now have a pretrained foundation model that can be fine-tuned on small labeled datasets - a practical advantage when ground-truth annotations for lunar features are limited. The open release on Hugging Face and GitHub means teams can replicate results and adapt the model for their own mapping, resource identification, or surface analysis projects without negotiating data access agreements. For researchers exploring how AI intersects with scientific workflows, AI Scientific Research Courses offer structured paths to build these exact skills - moving from understanding foundation models to applying them in domain-specific contexts.