The Pulse
Claude Science Fills a Third of the Sky in New UV Map
Anthropic researcher Brice Ménard used Claude Science to combine ultraviolet observations and estimate the unobserved portions of a full-sky map. Anthropic says the map labels measured and predicted regions and includes uncertainty estimate
AI.info Team · Written with AI ·
“I think it’s beautiful—and it can join the other canonical maps in helping generations of students appreciate the variety of physical processes at work in our galaxy.”
Brice Ménard, astrophysicist at Johns Hopkins University and researcher at Anthropic
Anthropic researcher Brice Ménard says he worked with Claude Science to produce what the company describes as the first complete map of the sky in ultraviolet light. Anthropic published his account on October 8, 2026. The map combines telescope measurements with computer-generated estimates: about one-third of it, including much of the Milky Way’s galactic plane, was predicted rather than directly observed in ultraviolet.
That distinction matters. The project does not add new telescope observations; it joins existing surveys, estimates the missing regions and marks pixels as measured or predicted, with uncertainty estimates available in additional map layers. Anthropic presents the result chiefly as an educational resource, meant to show students how the sky’s appearance changes across wavelengths.
Why ultraviolet maps have holes
Earth’s ozone layer absorbs ultraviolet light, so astronomers need space-based instruments to observe it. NASA’s GALEX mission, which operated from 2003 to 2013, supplied the largest dataset, imaging about two-thirds of the sky across roughly 38,000 observations, according to Ménard’s account. GALEX deliberately avoided areas around very bright stars, including parts of the Milky Way’s densely populated plane, to protect its detectors.
NASA’s Swift and South Korea’s FIMS/SPEAR missions collected additional ultraviolet data, but the combined surveys still left gaps. Their observations also came from different instruments and conditions, so the images could not simply be laid together. The team had to make each survey internally consistent, remove glare around bright stars, then cross-calibrate the datasets and align them to a common resolution and coordinate system.
Claude’s estimates fill the unobserved third
For the missing regions, Ménard asked Claude Science to use inpainting, a method for estimating absent parts of an image from surrounding information. The system compared the two-thirds of the sky with ultraviolet coverage against observations in visible, infrared and radio wavelengths, learning how ultraviolet brightness corresponded to those other signals. It then used those relationships to estimate ultraviolet values where no ultraviolet measurements existed.
Ménard describes a holdout test: Claude hid portions of regions that already had ultraviolet measurements and tried to reconstruct them without access to the concealed values. After several rounds of refinement, the estimates came within about 10% of the real measurements, he writes. That figure reflects the test on hidden, previously observed regions; it is not a direct measurement of accuracy across the sky’s genuinely unobserved areas.
Claude also added estimated ultraviolet light from more than 100 million individual stars, inferred from visible-light measurements by the European Space Agency’s Gaia satellite. The final product therefore combines measured diffuse ultraviolet light, modeled gaps and estimates for individual stars—not a uniform set of direct ultraviolet readings.
A missed artifact survived two review rounds
The process included a visible failure. Ménard noticed faint circular patterns in a dim region of the map: remnants of individual GALEX observation footprints, each carrying a slight glow from Earth’s atmosphere. Claude had identified leftover atmospheric glow as a possible issue, but the circles remained after two rounds of review by other agents.
After Ménard pointed them out, Claude traced the artifact to that residual glow and corrected it across all 38,000 GALEX observations. Anthropic says the work produced more than a dozen successive map versions over several days, with Claude handling hours of computation between planning exchanges.
The episode shows both the practical value and the limits of the account’s central claim. Claude coordinated data processing and helped make a long-deferred visualization feasible, but the result depended on a researcher’s judgment to spot a defect the automated reviews missed. Its clearest use, as described by Anthropic, is to help learners see structures such as dust clouds around young stars and large rings left by stellar explosions—while keeping predicted sky distinct from observed sky.