科學 AI 與開放模型
NASA and IBM Release Multimodal Lunar Foundation Model, Aligning Data from Nine Instruments into a Shared Representation
The NASA-IBM Lunar Foundation Model integrates more than 30 spatially aligned data layers from nine instruments across four missions and releases its weights under Apache 2.0. Official downstream experiments report reductions of up to 22% in ice-prediction error, but the results still come from the development team and have not yet been independently reproduced.
NASA and IBM have released the NASA-IBM Lunar Foundation Model. Rather than directly answering natural-language questions, the model is designed to encode lunar remote-sensing data—with widely varying resolutions, observation angles, and physical quantities—into a shared representation that can be fine-tuned for downstream tasks. The accompanying machine-learning dataset contains more than 30 spatially aligned layers sourced from nine instruments across four missions, including NASA’s Lunar Reconnaissance Orbiter, GRAIL, Lunar Prospector, and Japan’s SELENE/Kaguya. The [NASA project page](https://science.nasa.gov/artificial-intelligence-science/) confirms that the model covers both high-resolution imagery and geophysical data. The [Hugging Face model card](https://huggingface.co/nasa-ibm-ai4science/NASA-IBM-Lunar-Foundation-Model) lists an Apache 2.0 license and provides TerraTorch loading instructions along with three downstream checkpoints.
The architecture is derived from TerraMind, originally developed by IBM and ESA for Earth observation. It uses masked-token modeling to learn cross-modal relationships while handling data that ranges from surface imagery at roughly one meter per pixel to gravity measurements at approximately 20 kilometers per pixel. Downstream adaptation uses LoRA, with about 90% of the foundation model’s weights frozen, for three task categories: crater detection, irregular mare patch segmentation, and estimation of ice-resource potential in the Moon’s polar regions. This allows researchers to incorporate new spectral, topographic, or thermal data without retraining the entire backbone.
According to [release materials](https://newsroom.ibm.com/2026-09-10-ibm-and-nasa-release-open-source-ai-model-to-support-lunar-exploration) from the IBM and NASA team, the model reduced RMSE by up to 22% on the ice-potential task relative to an ImageNet-pretrained SwinV2-B. For crater detection at a scale of about 100 meters, it achieved an improvement of nearly 19% while using half the training data, while volcanic-terrain segmentation improved by roughly 3%. These percentages correspond to different datasets, metrics, and resolutions and cannot be combined into a general claim that the model is “23% more accurate.” The report, weights, data, and downstream models are now public, but the evaluation was still conducted by the same team that developed the system. Engineering and research teams should next examine whether cross-mission data splits prevent spatial leakage, how robust the model is when data from particular sensors is missing, and whether the uncertainty of its outputs is adequate for high-stakes decisions such as landing-site selection and resource prospecting.