IBM and NASA release open-source AI model for lunar exploration

Overview

Updated Sep 11

NASA is targeting 2028 for humans to walk on the Moon again. Before astronauts land, scientists must find ice in shadowed craters and identify safe terrain. On September 11, that mapping work got a new tool: an open-source AI model built by IBM and NASA.

The NASA-IBM Lunar Foundation Model is trained on more than 30 layers of data from nine instruments across four missions, including NASA's Lunar Reconnaissance Orbiter and Japan's SELENE/Kaguya spacecraft. It cuts error in identifying potential ice deposits by 22% versus a leading computer-vision baseline. The model and a co-registered lunar dataset are free to download from Hugging Face.

Why it matters

This shared lunar AI model gives scientists worldwide a free tool for mapping ice and craters, supporting NASA's 2028 crewed return.

Questions about this story

Free account needed to ask — your question is kept and asked for you right after sign-up. Answers are public.

No questions yet — be the first to ask.

Key Indicators

22%

Error reduction in lunar ice identification

Comparison of the NASA-IBM model to the SwinV2-B computer-vision baseline.

Missions contributing lunar data

Nine instruments across NASA's LRO and GRAIL missions plus JAXA's SELENE/Kaguya.

30+

Spatially-aligned data layers

Covering surface and subsurface properties of the Moon in a unified framework.

2M+

Co-registered data points in open dataset

Aggregated from tens of thousands of lunar images and maps.

Voices

Curated perspectives — historical figures and your fellow readers.

People Involved

Juan Bernabé-Moreno

Director, IBM Research Europe, UK and Ireland

Leading IBM's team on the Lunar Foundation Model

Organizations Involved

International Business Machines Corporation (IBM)

Technology company

Co-developer and publisher of the Lunar Foundation Model

Built the Lunar Foundation Model with NASA and released it open-source.

National Aeronautics and Space Administration (NASA)

Federal space agency

Co-developer and primary mission owner

Provided lunar data and domain expertise; the model supports the Artemis program's 2028 crewed return.

Japan Aerospace Exploration Agency (JAXA)

National space agency

Data contributor

Contributed SELENE/Kaguya mission data to the model's training set.

Hugging Face

AI platform company

Model host

Platform hosting the open-source Lunar Foundation Model and dataset.

Timeline

August 2026 September 2026

2 events Latest: September 11th, 2026 · 2 months ago

September 2026
IBM and NASA release open-source Lunar Foundation Model
Latest Announcement

Model and co-registered lunar dataset published on Hugging Face. Ice detection error cut by 22% versus baseline.
September 11th, 2026
August 2026
SpaceX Falcon 9 crashes into the Moon
Incident

The rocket's upper stage created a new crater. IBM later used the image to validate the model's crater detection.
August 5th, 2026

Scenarios

Lunar Foundation Model becomes standard tool for lunar science

Likely Resolves by Sep 11, 2027

Discussed by: IBM and NASA press materials highlight the model's crater mapping and ice detection for the Artemis program.

The open-source release and co-registered dataset become reference tools for the lunar science community. Researchers fine-tune the model for new tasks, and peer-reviewed lunar papers cite it as their analytical backbone. The model's crater and ice maps inform Artemis landing site selection.

Competing lunar foundation models fragment the field

Possible Resolves by Sep 11, 2027

Discussed by: Pattern of competing open-source foundation models in other AI domains, such as large language models.

Other research groups and companies train lunar foundation models on different data combinations and release them open-source. The community splits across multiple tools, preventing any single model from becoming standard. This mirrors what happened with general-purpose language models.

Model helps confirm lunar ice deposit for Artemis planning

Uncertain Resolves by Q2 2028

Discussed by: NASA's Artemis program goals and the model's 22% error reduction in ice detection.

The model's predictions guide targeted observations that confirm subsurface ice in a permanently shadowed polar crater. NASA cites the finding in Artemis planning for a water-producing lunar base.

Historical Context

3 moments from history that rhyme with this story — and how they unfolded.

August 2023

IBM's Prithvi EO Earth observation model (2023)

IBM released Prithvi EO, an open-source geospatial foundation model trained on NASA's Harmonized Landsat-Sentinel data. The model can be fine-tuned for tasks like flood detection and crop classification without retraining from scratch. It was the first geospatial foundation model deployed in orbit.

Then

It established IBM's pattern of building open, reusable scientific AI models and became the basis of the Prithvi family.

Now

The Prithvi family now spans Earth observation, weather, heliophysics, and the Moon.

Why this matters now

The Lunar Foundation Model is the direct descendant of Prithvi EO, using the same foundation-model approach and low-rank adapter fine-tuning techniques.

July 2021

DeepMind's AlphaFold open-source release (2021)

DeepMind released AlphaFold, an AI system that predicts protein structures from amino acid sequences, along with its source code and a database of over 350,000 predicted protein structures. Researchers worldwide adopted it for drug discovery, disease research, and more.

Then

The scientific community embraced it rapidly, with thousands of researchers using it within months.

Now

AlphaFold became the standard tool in structural biology, cited in tens of thousands of papers.

Why this matters now

AlphaFold showed that an open-source AI model can become the shared analytical foundation of an entire scientific field, replacing bespoke task-specific systems.

1961-1972

Apollo-era lunar mapping (1960s-1970s)

NASA scientists and cartographers mapped the Moon's surface manually using telescopic and orbital photography. The process produced charts that guided every Apollo landing, but it was painstaking and the maps had gaps, particularly at the poles.

Then

Apollo missions landed successfully with the maps available at the time.

Now

The lunar community spent decades filling map gaps with later missions like Lunar Prospector and the Lunar Reconnaissance Orbiter.

Why this matters now

The Lunar Foundation Model automates this mapping tradition, covering the entire Moon with co-registered data from four missions in a single unified model.