Claude Science fills in the universe's missing UV map

Ask any astronomer to show you the sky, and you will usually get a tour across multiple wavelengths. Visible light shows the stars. Infrared reveals glowing dust grains along the galactic plane. Radio picks up cold hydrogen, and X-rays capture violent black holes and exploding stars.

There has always been an embarrassing hole in this collection: ultraviolet light. Earth's ozone layer blocks UV rays, meaning telescopes must head into space. While missions like NASA’s GALEX mapped roughly two-thirds of the sky over a decade, it intentionally skipped the galactic plane—the crowded highway of our galaxy—to avoid burning out its delicate sensors.

Now, Johns Hopkins astrophysicist and Anthropic researcher Brice Ménard has fixed the map using Claude Science. By orchestrating a crew of AI agents, they produced the first complete full-sky ultraviolet map, predicting the missing third with striking accuracy.

Delegating the grunt work to AI agents

Piecing together a global map sounds simple until you look at the calibration required. You have to locate public surveys, harmonize coordinate systems, strip away stellar glare, and align mismatched instruments. For human researchers, pixel-level calibrations take weeks of monotonous work that rarely justifies derailing higher-priority physics.

Ménard used Claude Science as an automated project manager. The AI deployed parallel agents to scrape and combine catalogs from NASA’s GALEX, Swift, South Korea’s FIMS/SPEAR, and the European Space Agency’s Gaia satellite.

The process hit snags. Early on, Ménard noticed faint, repeating discs across some of the dimmest sky patches. Each GALEX observation covers a circular field of view, carrying a subtle glow from Earth's upper atmosphere. The AI review agents had missed the pattern twice. Once Ménard flagged it, Claude tracked down the leftover atmospheric glow and recalculated corrections across all 38,000 individual GALEX frames in a few hours.

Reconstructing a third of the sky

The toughest nut to crack was the unobserved third of the sky. Instead of inventing textures out of thin air, Claude utilized multi-wavelength inpainting.

While UV telescopes had skipped these blind spots, astronomers already have extensive observations of them in visible, infrared, and radio bands. Claude examined the two-thirds of the sky with complete coverage to learn how UV brightness correlates with other wavelengths. It then applied those learned relationships to the blind spots, generating pixel-level UV estimates along with uncertainty scores.

To test the math, the team deliberately masked out known patches and had the model reconstruct them blind. Claude's inpainted predictions landed within 10% of ground-truth measurements—differences nearly invisible to the human eye. Finally, the model mapped inferred UV light for over 100 million individual stars using Gaia's visible-light data, revealing massive stellar explosion rings, luminous dust clouds, and the neighboring Magellanic Clouds.

Why it matters

For scientists and data workers, this project shows how to rescue "backlog science." Almost every discipline has datasets that desperately need merging or cleaning, but no lab can justify dedicating a postdoctoral researcher to months of repetitive wrangling. Ménard’s workflow proved that high-level human oversight combined with parallel AI execution can turn tedious side quests into finished tools without stealing time from primary research.

For the AI industry, it is a clean proof of concept for autonomous agent orchestration. The tool did not just chat; it retrieved files, normalized heterogeneous sensors, ran complex multi-wavelength regression, and accepted real-time critical feedback to rerun massive batch jobs.

Still, we think this breakthrough highlights a lingering risk: AI confirmation bias. The subtle GALEX observation footprints slipped straight through two independent layers of AI peer review and were only spotted when a veteran human looked closely at the dimmest pixels. Machine learning can model the cosmos with jaw-dropping precision, but until AI develops genuine skepticism of its own artifacts, high-stakes science still demands a human holding the magnifying glass.