The First Complete UV Sky Map: How Claude Science Filled 50 Years of Observation Gaps

How the first complete UV sky map came to be

Brice Ménard, an astrophysicist at Johns Hopkins University and an Anthropic researcher, describes creating the first-ever complete sky map in ultraviolet light. The project, done jointly with Claude Science, combines two wavelengths: far-UV (154 nm) and near-UV (232 nm). About a third of the map was predicted using AI, while the rest was assembled from every available UV observation collected over the past 50 years.

Ménard teaches astrophysics and often shows students how the sky looks completely different depending on the wavelength: in visible light you see stars, in infrared you see glowing dust clouds, in radio you see hydrogen gas (the raw material for new stars), and in X-ray you see stellar explosions and black holes. Together, these views reveal how gas collapses into stars, how stars spread dust across a galaxy, and how galaxies like ours form.

Why there was no complete UV map for 50 years

UV radiation is absorbed by Earth's ozone layer, so only a space telescope can observe the cosmos in the UV range. Over the past 50 years, several space observatories captured pieces of the sky and published their data, but a complete picture never came together.

The largest database is NASA GALEX (launched in 2003, retired in 2013). Over 10 years, it made roughly two-thirds of all sky observations: 38,000 individual images, collected under varying conditions. But GALEX deliberately avoided regions with bright stars, especially the plane of the Milky Way (where stars are most densely packed), because these regions risked damaging the satellite's sensitive detectors.

Other space telescopes added to the database — NASA Swift and South Korea's FIMS/SPEAR — but even combined, their data could only close part of the gaps. Building a complete map required manually running statistical techniques, calibrating millions of pixel values, and redoing the analysis multiple times. That takes weeks of painstaking work. No astrophysicist considered this work a priority — it matters for teaching students, but not for making new discoveries.

How Claude Science built the map in days instead of months

In the summer of 2026, Ménard gave Claude Science a simple-sounding task: gather every available UV dataset, bring them to a common scale, merge them into one map, and fill in all the gaps. Simple to state, difficult to execute.

Claude deployed a team of AI agents that:

• Searched the internet for every publicly available UV sky survey (for GALEX alone, this meant tens of thousands of images) • Standardized each dataset so that photos taken in different years and under different conditions could be compared • Removed interference from bright stars (which added "glare" to the images, much like a streetlight causes flare in an earthly photo) • Cross-calibrated the different telescopes against each other: NASA GALEX, NASA Swift, and FIMS/SPEAR each capture UV light slightly differently, so everything needed to be mapped onto a single coordinate grid

Then came the hard part: filling the gaps. A third of the sky had never been observed in UV. Ménard asked Claude to use inpainting — a machine-learning method in which a model trains on millions of images and learns to reconstruct damaged or missing areas.

Claude combined this with additional information: wherever UV data was missing, observations existed at other wavelengths (visible, infrared, radio). Across the two-thirds of the sky where UV data already existed, Claude learned the correlation between UV brightness and brightness at other wavelengths. It then applied that pattern to the UV-free third of the sky and predicted what it should look like.

To check accuracy, Ménard asked Claude to take regions with known UV data, hide parts of them, and ask the model to reconstruct what was hidden. After several rounds of refinement, accuracy reached within 10% of the real UV measurements — a difference imperceptible to the human eye.

Finally, Claude added UV light estimates for more than 100 million individual stars, inferred from visible-light observations by ESA's Gaia mission.

Refinement and the final result

The first version of the map was nearly perfect, but Ménard spotted artifacts: phantom rings around certain areas of the sky — fingerprints left by individual GALEX observations. Each GALEX image captures a circular patch of sky with residual traces of Earth's atmospheric glow. Claude missed this in two rounds of review, but once Ménard pointed it out, the agents traced the problem to its source and reprocessed all 38,000 observations. Within a couple of hours, the artifacts were gone.

The final result: a UV sky map revealing dust clouds around young stars, giant dusty rings and loops (traces of massive stellar explosions), faint filaments of dust outside the galactic plane, and the Large and Small Magellanic Clouds. The map became a worthy addition to other canonical views of the sky and turned into a powerful educational tool. For Ménard personally, it was the first time he managed to complete a significant project without taking time away from his core research work.