Exploring the Moon requires navigating an overwhelming ocean of scientific data. Over the past five decades, space probes, orbiters, and surface landers have beamed back petabytes of high-resolution imagery, radar scans, gravity maps, and thermal readings. However, turning these scattered data archives into actionable landing maps and resource guides has historically forced planetary scientists to sift through images manually or rely on narrow, low-resolution software tools.
To overcome this data bottleneck, the National Aeronautics and Space Administration and computing giant IBM have released the NASA-IBM Lunar Foundation Model. Published as an open-source tool on Hugging Face with full codebase access on GitHub, the artificial intelligence model gives researchers worldwide a unified platform to analyze the lunar surface at scale. Trained on decades of multi-mission data, the foundation model accurately identifies potential water ice deposits, maps hazardous impact craters, and pinpoints ancient volcanic features. The open-source system provides space agencies and private aerospace companies with the digital navigation tools required to establish a sustained human presence on the Moon under the Artemis program.
Overcoming Decades of Fragmented Lunar Observation Data
The primary challenge in lunar science has never been a lack of observations; it has been the difficulty of integrating incompatible data formats gathered across different space missions.
Unifying 30 Layers of Multi-Instrument Mission Data
To train the foundation model, research teams from IBM and NASA constructed the first unified, machine-learning-ready planetary dataset of its kind. The dataset aggregates more than 30 spatially aligned layers of geophysical information collected by nine scientific instruments operating across four distinct missions.
Rather than looking at single photographs in isolation, the foundation model processes multiple data streams simultaneously. It correlates high-resolution optical camera feeds with laser altimeter topography, neutron spectrometer water-hydrogen signatures, thermal radiometer heat maps, and gravity anomaly measurements. By merging these distinct sensory layers into a cohesive spatial framework, the model uncovers complex geological relationships that remain invisible when scientists examine single data layers on their own.
Processing Petabytes from the Lunar Reconnaissance Orbiter and GRAIL
The core training dataset relies heavily on data collected by NASA’s Lunar Reconnaissance Orbiter, which has orbited the Moon for 17 years. The mission generated more data than all other NASA planetary exploration missions combined, capturing an almost seamless, high-resolution mosaic of the entire lunar surface.
The foundation model digested roughly 2 million individual image tiles from the orbiter’s archive. This massive training library includes more than 1 million optical camera photographs resolving details down to 1 meter per pixel, alongside nearly 964,000 multispectral images captured at 100-meter resolution.
The researchers combined this imagery with sub-surface gravity field data from NASA’s Gravity Recovery and Interior Laboratory mission, magnetic field surveys from Lunar Prospector, and complementary terrain scans from the Japan Aerospace Exploration Agency’s SELENE Kaguya orbiter. Training a large neural network on this diverse multi-agency archive created a general-purpose model that understands lunar geology across multiple spatial resolutions and lighting conditions.
Benchmarking Accuracy in Lunar Ice and Crater Detection
Unlike traditional machine learning models that require engineers to build and train specialized algorithms from scratch for a single narrow task, foundation models acquire broad general knowledge through pre-training on unlabeled datasets. Scientists can then fine-tune the model for specific research goals using only small amounts of labeled data.
A 23 Percent Reduction in Mapping Errors
In rigorous benchmark evaluations, the NASA-IBM Lunar Foundation Model demonstrated superior accuracy compared to established computer vision systems. When tested against standard image analysis models, including Microsoft’s SwinV2-B vision architecture, the new lunar foundation model identified key geographic surface features up to 23% more accurately.
The model showed notable performance gains in crater detection and classification tasks. At a spatial resolution of approximately 100 meters per pixel, the foundation model improved crater identification accuracy by nearly 19% while requiring only half the amount of labeled training data used by conventional vision systems.
The model proved its real-world capability during a live test involving an unexpected surface event: when a spent SpaceX Falcon 9 rocket upper stage impacted the lunar surface, IBM researchers fed post-impact orbital imagery into the system. The model correctly identified the fresh impact site as a newly formed crater on its first attempt, successfully distinguishing the new debris mark despite its overlap with an older, pre-existing crater.
Spotting Subsurface Ice in Permanently Shadowed Regions
The most strategically vital application of the foundation model is finding water ice deposits near the lunar poles. The Moon’s rotational axis has a tilt of only 1.5 degrees, meaning deep craters at the lunar south pole never receive direct sunlight. These permanently shadowed regions function as ultra-cold thermal traps where temperatures plunge below minus 240 degrees Celsius, preserving frozen water ice for billions of years.
Observing permanently shadowed regions from orbit is exceptionally difficult because standard optical cameras see only total darkness. The foundation model overcomes this visual barrier by combining laser topography, surface roughness data, secondary reflected sunlight, and subsurface thermal stability models. In benchmark tests targeting high-potential ice zones, the model reduced spatial mapping errors by 22%. By pinpointing where subsurface ice patches remain thermally stable within a few feet of the lunar regolith, the model allows mission planners to target exact coordinates for robotic drill sampling.
Practical Applications for the Artemis Program and Lunar Bases
The open-source release of the lunar foundation model comes at a decisive moment for international space exploration. NASA’s Artemis campaign plans to land astronauts near the lunar south pole and construct permanent surface habitats before the end of the decade.
Selecting Safe Landing Sites and Navigating Hazardous Terrain
Landing a crewed spacecraft on the Moon is a dangerous engineering task. The lunar south pole features rugged terrain, dramatic mountain peaks, steep crater walls, and long, deceptive shadows that wash out visual depth perception. A lander touching down on a slope steeper than 15 degrees or striking a boulder larger than 1 meter runs the risk of tipping over.
The NASA-IBM foundation model automates terrain risk assessment across expansive potential landing corridors. The software analyzes slope angles, crater density, boulder distribution, and soil mechanics across prospective landing zones in seconds. Flight dynamics teams can use the model to evaluate multiple candidate sites simultaneously, identifying smooth, obstruction-free landing zones that sit close to scientific points of interest while keeping astronauts safe from terrain hazards.
In-Situ Resource Utilization for Rocket Fuel and Oxygen
A sustainable, permanent human presence on the Moon depends on in-situ resource utilization—the practice of harvesting raw materials directly from the lunar environment rather than hauling every pound of supplies from Earth. Hauling water into deep space costs tens of thousands of dollars per kilogram, making long-term supply flights from Earth financially unsustainable.
Lunar water ice represents a vital resource for space exploration. Using solar or nuclear power systems, future lunar bases can melt the ice, purify it into drinking water, and split the water molecules through electrolysis into breathable oxygen and liquid hydrogen. Liquid hydrogen and liquid oxygen serve as the primary chemical propellants for deep-space rockets.
By mapping the depth, concentration, and accessibility of polar ice reserves, the foundation model helps space agencies design commercial water-mining infrastructure. Establishing refueling depots on the Moon will allow spacecraft to refuel on the lunar surface, lowering the cost of launching human exploration missions to Mars and beyond.
The Expansion of Open-Source Geospatial Foundation Models
The lunar artificial intelligence model represents the third major milestone in an ongoing scientific partnership between NASA and IBM Research to democratize access to advanced space data.
From Prithvi Earth Observation to Surya Solar Weather
The collaborative initiative began with the development of the Prithvi family of open-source foundation models. The first Prithvi model focused on terrestrial geospatial analysis, utilizing petabytes of satellite imagery from the harmonized Landsat and Sentinel-2 archives. Environmental scientists, urban planners, and agricultural researchers use Prithvi to track deforestation, map flood inundation, detect crop health anomalies, and identify wildfire burn scars worldwide.
Following the success of terrestrial geospatial AI, the partnership launched Surya, the first open-source heliophysics foundation model trained on high-resolution observations from NASA’s Solar Dynamics Observatory. Surya analyzes the Sun’s dynamic magnetic field, predicting solar flares, coronal mass ejections, and geomagnetic storms that can disable electrical power grids, disrupt satellite communications, and expose astronauts to dangerous radiation bursts. The release of the lunar foundation model extends this open-source lineage from Earth and the Sun directly into planetary exploration.
Democratizing Planetary Exploration for Global Researchers
Historically, planetary science required access to specialized supercomputing facilities and proprietary laboratory software that was available only to well-funded universities and national space agencies. Independent researchers, small commercial space startups, and students in developing nations faced steep barriers to entry.
By hosting the model weights and codebase on public open-source repositories, NASA and IBM are leveling the scientific playing field. Any university research laboratory or commercial robotics startup can download the foundation model, fine-tune it on a local workstation, and deploy custom mapping algorithms for their own lunar rovers or landers. This open approach encourages global scientific collaboration, accelerating the pace of lunar discoveries by allowing thousands of independent programmers and geologists to build customized analytical tools on top of a shared, verified AI foundation.
Long-Term Outlook for Autonomous Planetary Science
The deployment of foundation models to study the Moon marks a fundamental transition in how robotic spacecraft and human explorers interact with deep-space environments.
Scaling Foundation Models Beyond the Moon to Mars and Asteroids
The underlying architecture of the NASA-IBM model provides a reusable blueprint for robotic exploration across the entire solar system. Space agencies have accumulated massive, multi-instrument archives from exploration missions targeting Mars, Venus, Mercury, and the outer moons of Jupiter and Saturn.
NASA’s Science Mission Directorate plans to adapt the foundation model framework to study other celestial bodies. Future iterations will process multi-spectral imagery and radar data from Mars orbiters to map underground glaciers, assess flash-flood channels, and locate safe landing sites for crewed missions to the Red Planet. Similar models will analyze radar scans of Saturn’s moon Titan and gravity data from metallic asteroids, automating geological classification across worlds that lie millions of miles from Earth.
Accelerating Human Lunar Settlements Through Digital Discovery
The establishment of permanent lunar infrastructure—including radio astronomy observatories on the lunar far side, pressurized habitats, and solar power stations on permanently lit crater rims—requires precise geological insight.
As commercial companies like SpaceX, Blue Origin, and Intuitive Machines prepare commercial cargo landers, the demand for precise lunar mapping data will expand exponentially. The NASA-IBM Lunar Foundation Model turns historical space archives into an active discovery engine. By identifying mineral distributions, mapping volcanic structures known as Irregular Mare Patches, and tracking micro-meteorite impact rates, the software helps engineers design durable surface structures that can withstand the harsh lunar environment for decades.
A New Era of Open Space Intelligence
The launch of the NASA-IBM Lunar Foundation Model represents a major leap forward in planetary exploration. By combining 17 years of lunar observations with modern neural network architectures, NASA and IBM have created an open-source tool that turns raw scientific data into actionable exploration maps.
From detecting buried water ice deposits in permanently shadowed polar craters to identifying smooth landing zones for the Artemis astronauts, the foundation model provides the navigation and resource intelligence required to build a permanent human presence on the Moon. By releasing the software openly to the global scientific community, the creators ensure that the exploration of space remains an inclusive, collaborative journey. As humanity prepares to return to the lunar surface, open-source artificial intelligence is illuminating the path forward into the deep-space frontier.
