Claude Science Builds First Complete Ultraviolet Sky Map With 10% Error Margin

Anthropic has unveiled the first complete map of the sky in ultraviolet light, created by artificial intelligence. The project, called Claude Science, is meant to show how AI can take over the kind of tedious, routine work that researchers keep pushing to the back burner.

Astrophysicist Brice Ménard of Johns Hopkins University described the project. According to him, ultraviolet light reveals dust, gas clouds around young stars, rings left behind by stellar explosions, and other structures invisible in other parts of the spectrum.

Why a complete map didn't exist before

A complete UV sky map didn't exist until now for a simple reason: Earth's ozone layer blocks ultraviolet radiation, so UV observations are only possible from space. NASA's GALEX mission covered roughly two-thirds of the sky, but it deliberately skipped bright star-forming regions because of technical limitations of the instrument.

How AI pieced together the full map

To build the complete map, Claude Science used AI agents that carried out a series of sequential operations. The AI downloaded data from multiple space missions, calibrated that data to bring it to a common standard, and then merged the calibrated information into a single unified map.

The critical step was reconstructing missing sections using inpainting. This is a technique in which a neural network model learns from existing data and uses the patterns it finds to reconstruct the gaps. In testing, the average deviation of the predictions came out to about 10% from actual measurements — a level of accuracy sufficient for scientific use.

Practical significance

The project's creators are positioning the complete UV sky map as a teaching resource for the scientific community. Ménard suggested the work could inspire many researchers — plenty of scientists have likely shelved similarly large-scale data processing and analysis projects, deeming them too labor-intensive. Thanks to the growth of AI agents and automated workflows, such projects are now becoming achievable in relatively short timeframes.

The Claude Science project demonstrates AI's practical potential not in creating something fundamentally new, but in automating work with large volumes of existing data — work that is necessary but often gets deferred as too routine.