IBM and NASA Release the Lunar Foundation Model on 10 September 2026
On 10 September 2026, IBM and NASA jointly released the NASA-IBM Lunar Foundation Model as an open-source AI model, hosted on Hugging Face and with the complete codebase published on GitHub. The model is among the first open-source foundation models built specifically for lunar science and is the latest addition to IBM's Prithvi family of scientific foundation models, which previously spanned geospatial, weather, and heliophysics domains. The Lunar Foundation Model was trained on approximately two million image tiles of the Moon's surface from NASA's Lunar Reconnaissance Orbiter — more than one million high-resolution camera images and almost 964,000 multispectral images. IBM and NASA also released SomBench, a curated evaluation dataset for standardised lunar remote-sensing tasks, alongside the model. On SomBench, the Lunar Foundation Model exceeds widely used methods by up to 23 percent at identifying key geographic features on the lunar surface.
The Three Features the Model Detects
The Lunar Foundation Model is designed to identify three categories of scientifically significant surface features from LRO remote-sensing data. The first is potential water ice deposits in permanently shadowed regions near the lunar poles. Identifying ice location and concentration is a prerequisite for evaluating potential sites for sustained human presence on the Moon, since water ice can serve as drinking water and, through electrolysis, as a rocket propellant feedstock. The second category is crater formations — detecting crater boundaries, estimating depths, and mapping spatial distributions across the surface. The third is volcanic features, including ancient lava plains and dome structures left by the Moon's geological past. On the ice-identification task specifically, the model pinpoints potential lunar ice deposits 22 percent more accurately than prior widely used methods. The 23 percent error reduction holds across the broader set of SomBench evaluation tasks covering all three feature categories.
The IBM Prithvi Family of Scientific AI Models
The Lunar Foundation Model is the most recent member of IBM's Prithvi family of open foundation models for scientific and environmental domains. Prithvi Geospatial was the first in the family, trained on NASA's Harmonised Landsat and Sentinel-2 satellite imagery to support Earth observation tasks including flood mapping, crop boundary segmentation, and wildfire extent detection. Prithvi Weather applied the methodology to atmospheric data for medium-range forecasting. Prithvi Heliophysics covers solar and space weather data. The consistent pattern across the family is pre-training on large volumes of unlabelled domain-specific observational data to build transferable representations, which researchers then fine-tune for specific downstream tasks using a much smaller labelled dataset. This approach is particularly effective in scientific domains where labelled examples are scarce but observational data from instruments running continuously over decades is abundant.
Why an Open-Source Foundation Model Changes Lunar Research
Before the Lunar Foundation Model, teams conducting remote-sensing analysis of the Moon had to train separate task-specific models from scratch for each analytical objective — crater counting, ice mapping, terrain classification — requiring access to the full LRO archive and the computational resources to process it. This bottleneck disadvantaged independent research groups, university teams, and national space agencies without large AI compute infrastructure. An open-source foundation model hosted on Hugging Face changes this dynamic. A research team can fine-tune the Lunar Foundation Model on a modest labelled dataset specific to their mission target without rebuilding the full LRO training pipeline. The Hugging Face distribution makes the model accessible through standard library imports, and the GitHub codebase provides documented entry points for inference and fine-tuning without bespoke data engineering.
What Lunar AI Means for Indian Space and Geospatial Teams
India's space programme is an active lunar participant. The Indian Space Research Organisation's Chandrayaan missions have produced their own lunar surface datasets, and ISRO holds remote-sensing imagery that can serve as fine-tuning data for mission-specific applications. For ISRO researchers and academic institutions working with Chandrayaan instrument data, the Lunar Foundation Model provides a starting point for producing India-calibrated terrain classifiers faster than training from scratch. Beyond planetary science, the Prithvi pre-training methodology — building representations from large unlabelled satellite datasets — transfers directly to Indian Earth-observation challenges: agricultural land-use mapping from Resourcesat imagery, flood extent detection from Cartosat, and urban growth tracking from IRS satellite data. The National Remote Sensing Centre, state disaster management authorities, and agricultural research organisations all work with exactly the kinds of high-volume satellite imagery that Prithvi-style models are designed to leverage.
The Bottom Line
On 10 September 2026, IBM and NASA released the NASA-IBM Lunar Foundation Model as an open-source model on Hugging Face, with the full codebase on GitHub, as part of IBM's Prithvi scientific foundation model family. The model was trained on approximately two million image tiles from NASA's Lunar Reconnaissance Orbiter — more than one million high-resolution camera images and nearly 964,000 multispectral images. On the SomBench benchmark released alongside it, the model outperforms widely used methods by up to 23 percent on lunar surface feature identification and locates potential ice deposits 22 percent more accurately. The open-source release removes the compute barrier for research teams fine-tuning on mission-specific data. For Indian space institutions and geospatial AI teams, the Prithvi pre-training methodology is directly applicable to India's own satellite imagery datasets across agricultural mapping, flood detection, and terrain analysis.
Frequently Asked Questions
What is the NASA-IBM Lunar Foundation Model and when was it released?+
The NASA-IBM Lunar Foundation Model is an open-source artificial intelligence model for lunar science research released on 10 September 2026 by IBM and NASA. It is part of IBM's Prithvi family of scientific foundation models and was trained on approximately two million image tiles from NASA's Lunar Reconnaissance Orbiter — more than one million high-resolution camera images and almost 964,000 multispectral images. The model is hosted on Hugging Face and the complete codebase is available on GitHub. IBM and NASA also released SomBench, a benchmark for evaluating lunar remote-sensing models on standardised tasks. On SomBench, the model outperforms widely used methods by up to 23 percent on identifying key geographic features and pinpoints potential lunar ice deposits 22 percent more accurately than prior methods.
What surface features can the Lunar Foundation Model identify?+
The Lunar Foundation Model is designed to identify three categories of scientifically significant lunar surface features from remote-sensing data. The first is potential water ice deposits in permanently shadowed polar regions — identifying ice location is critical for evaluating sustained human presence sites, since ice can be used as drinking water and as a rocket propellant feedstock through electrolysis. The second category is crater formations, covering boundary detection, depth estimation, and spatial distribution mapping. The third is volcanic features including lava plains and dome structures. The model achieves 22 percent greater accuracy than prior methods on the ice-identification task and a 23 percent error reduction across the full SomBench evaluation set. The model accepts LRO imagery as input, making it directly applicable to data collected by the Lunar Reconnaissance Orbiter's instruments since 2009.
What is IBM's Prithvi family of scientific foundation models?+
IBM's Prithvi family is a series of open foundation models for scientific and environmental domains, each pre-trained on large volumes of domain-specific observational data to build transferable representations for downstream scientific tasks. Prithvi Geospatial was the first, trained on NASA's Harmonised Landsat and Sentinel-2 satellite imagery for Earth observation tasks including flood mapping, crop segmentation, and wildfire boundary detection. Prithvi Weather applies foundation model training to atmospheric data for medium-range weather prediction. Prithvi Heliophysics covers solar and space weather domains. The Lunar Foundation Model extends the family to extraterrestrial remote sensing. All Prithvi models are available on Hugging Face, allowing researchers to fine-tune for their specific task using a small labelled dataset rather than training a model from scratch on the full observational archive.
How can Indian engineering and research teams use the NASA-IBM Lunar Foundation Model?+
Indian teams can use the NASA-IBM Lunar Foundation Model in two practical ways. For space science, ISRO and academic institutions working with Chandrayaan mission data can fine-tune the model on Chandrayaan imagery to produce India-calibrated lunar terrain classifiers without training from scratch. For Earth-observation and geospatial AI, the Prithvi pre-training methodology applies directly to Indian satellite imagery from Resourcesat, Cartosat, and IRS instruments. Teams at the National Remote Sensing Centre, state disaster management authorities, and agricultural research organisations can apply the same approach to flood mapping, agricultural land-use classification, and urban growth detection using their existing satellite archives. Model weights are accessible via the Hugging Face Python library, and the GitHub codebase provides standard fine-tuning entry points without requiring bespoke data pipeline infrastructure.
Written by
TechPillow Team
Sharing insights on technology, product development, and the Indian tech ecosystem.
