AI Fills Missing Third of Sky to Create First Complete Ultraviolet All-Sky Map

Background: The Missing Ultraviolet Sky

Ultraviolet light is absorbed by Earth's ozone layer, so UV observations require space telescopes. NASA's GALEX mission (2003–2013) conducted ~38,000 observations covering about two-thirds of the sky. However, GALEX deliberately avoided very bright stars and the Galactic plane to protect its detectors, leaving roughly one-third of the sky unobserved in UV.

AI-Driven Map Construction

Astrophysicist Brice Ménard (Johns Hopkins University, Anthropic) collaborated with Claude Science to build the first complete UV all-sky map. The workflow involved:

Data collection: AI agents searched and downloaded all publicly available UV survey datasets, each containing many individual images taken under different conditions.

Internal consistency: Agents calibrated each survey internally so images from different epochs could be compared.

Cross-calibration and merging: Because datasets came from different telescopes and instruments, agents cross-calibrated them, resampled to a common resolution, and projected onto a shared coordinate system.

Filling the Gaps with Inpainting

The hardest part was the ~33% of the sky never observed in UV. Ménard instructed Claude Science to use inpainting combined with multi-wavelength data (visible, infrared, radio) available for those regions. Using the two-thirds of the sky with UV measurements, the model learned the relationship between UV emission and other wavelengths, then applied that relationship to predict UV flux in the unobserved regions, along with uncertainty estimates.

Validation

To test accuracy, Ménard had Claude hide portions of the known UV data, create artificial gaps, and predict the missing values. After several iterations, the model achieved predictions within 10% of the true UV measurements — a difference nearly imperceptible to the human eye.

Adding Stellar Sources

On top of the inpainted diffuse background, Claude added UV estimates for over 100 million individual stars , inferred from the European Space Agency's Gaia visible-light measurements.

Iterative Refinement

Claude did not get everything right on the first attempt. Ménard and Claude produced over a dozen versions of the map, each more refined than the previous one, before arriving at the final version.

Resulting Map and Scientific Value

The final map (combining far-UV 154 nm and near-UV 232 nm) reveals numerous structures:

Dust clouds glowing around young, hot stars.

Giant dust rings and loops tracing bubbles blown by massive stellar explosions.

Faint dust filaments far from the Galactic plane, illuminated by the integrated starlight of the galaxy.

The Large and Small Magellanic Clouds, bright with young stars, below the Galactic plane.

Ménard notes the map joins the ranks of classic all-sky atlases, helping generations of students explore the universe.

AI for Science Workflow

The collaboration spanned several days with a rhythmic pattern: Ménard provided high-level guidance in a few interactions, then Claude executed hours of autonomous computation while Ménard worked on other projects. This demonstrates how AI can advance projects that scientists would otherwise defer due to time constraints.

Reference:

https://www.anthropic.com/research/the-missing-map-of-the-sky