TLDRocket
Sign in

Introducing IBM and NASA's new foundation model for the Moon

IBM Research

IBM and NASA opened a new AI model for mapping the Moon. It pulls old mission data into one map to spot ice, craters, and safer places to land.

Based on reporting by IBM Research — read the original for the full story.

Summary, retelling and take written by AI under human oversight; images are AI-generated illustrations. How we work · Report an error

IBM and NASA have put a new lunar foundation model into the open, and the pitch is straightforward: the Moon is messy, so give scientists one system that can digest the mess. The model, called NASA-IBM Lunar Foundation Model, pulls together decades of data from US and Japanese missions and is meant to be reused across tasks, not trained from scratch every time someone wants a new map.

That matters because lunar data comes in all shapes and resolutions. GRAIL measured the Moon’s gravity at 20 kilometers per pixel. LRO, by contrast, has worked at 1 meter per pixel, hunting for ice in dark polar craters and picking out tiny boulders and crater rims. IBM and NASA used a TerraMind-based architecture to make those different views line up, then fine-tuned it with lightweight LoRAs that left 90% of the base weights frozen.

NASA is already pointing the model at three jobs: finding smaller craters that still haven’t been catalogued, studying volcanic history, and searching polar craters for ice. The Moon’s lighting makes all of this harder than it sounds. A lunar day brings two weeks of sunlight and then two weeks of darkness, with no atmosphere to soften the glare or hide the shadows, and temperatures can swing from 250°F to -410°F depending on where the Sun lands.

The early results are solid enough to make the open-source release feel less like a demo and more like a useful tool. IBM and NASA say the model cut error rates by 22% on ice prospecting compared with a SwinV2 transformer, matched crater detection at one-meter resolution, and beat the same model by nearly 19% at 100 meters per pixel using half the training data. It also did better than a task-specific Swin model by 3% when mapping irregular mare patches, those odd volcanic features that may hold some of the Moon’s youngest rock.

That sits neatly inside NASA’s broader Artemis bet. The agency wants a base on the Moon, and if humans are really going to live there, they’ll need better maps than the ones built for flags and footprints. This one is for actual work: water, power, shelter, and not driving a rover into a crater because the shadows looked artistic from orbit.

My take — AI-written commentary, not fact-checked reporting

Open-sourcing this is the rare sensible move in space AI. Moon data is too valuable to leave trapped behind vendor theater, and the best way to find ice or bad terrain is to let other researchers kick the tires. The bigger trend is obvious: the winners won’t be the loudest model builders, but the teams that make old data useful again without pretending every problem needs a shiny new moonshot.

Read more about this at: IBM Research

Related stories

The daily briefing

Every AI story that matters, in your inbox by 8am.

TLDRocket reads all relevant sources, removes duplicate coverage, and summarises the day in two minutes. Follow companies and topics for alerts, or get the briefing in Slack. Free, no spam, unsubscribe anytime.