Astrophysicist uses Claude to complete a UV sky map—with a third predicted

The ultraviolet sky now has a complete map, though not a complete set of observations. In an October 8 account published by Anthropic, astrophysicist Brice Ménard describes using Claude Science to combine telescope surveys and predict the missing third. The result is intended for teaching, with layers distinguishing measured light from estimates and showing their uncertainty.

Why the UV sky had holes

Ménard, who works at Johns Hopkins University and Anthropic, calls the result the first complete ultraviolet sky map. UV observations require space telescopes because Earth’s ozone layer absorbs that light. NASA’s GALEX mission captured about two-thirds of the sky, but skipped very bright stars—including much of the Milky Way’s densely populated plane—to protect its detectors.

Claude coordinated agents to collect public UV surveys and reconcile images taken under different conditions. They removed glare near bright stars, aligned the surveys’ brightness scales, and put them at the same resolution and coordinates. This calibration workflow let measurements from different instruments become one map, rather than a patchwork of incompatible images.

Filling gaps with other kinds of light

For unobserved regions, Ménard asked Claude to use inpainting: estimating missing parts of an image from learned relationships. The system learned how UV brightness related to visible, infrared and radio observations in regions with known UV data. It then applied those relationships where UV measurements were absent, producing both brightness predictions and confidence estimates.

Ménard tested that approach by hiding regions with known UV measurements and asking the system to reconstruct them. After several rounds of refinement, he says the estimates came within about 10% of the real measurements. That result concerns hidden, already-observed regions—not direct verification of the sky’s unobserved third.

The final step added estimated UV light from more than 100 million individual stars, inferred from visible-light measurements by the European Space Agency’s Gaia satellite. Those stellar estimates sit on top of the reconstructed diffuse background.

The flaw two AI reviews missed

Human inspection caught a problem the agents had missed. Ménard noticed faint circles marking the boundaries of individual GALEX images. Residual atmospheric glow made those patches slightly brighter or darker than their neighbors. Claude had identified the risk at the outset, yet the map passed two agent reviews. After Ménard flagged it, Claude corrected all 38,000 observations; the circles disappeared after a couple of hours of processing.

The collaboration produced more than a dozen versions over several days. Ménard describes planning the next steps with Claude, then returning to other work while it computed for hours. The finished map depicts illuminated dust clouds, rings left by stellar explosions and faint filaments across the sky—features he wants students to explore alongside maps at other wavelengths.