A complete map of the sky in ultraviolet light has been assembled with help from Anthropic’s Claude Science. The project brought together data from multiple space missions and used AI agents to handle work that can be tedious for researchers: downloading, calibrating and combining observations.
Why ultraviolet observations need space missions
Ultraviolet light can reveal dust illuminated by starlight. The source describes examples including clouds around young stars and rings left behind by stellar explosions. Those features make ultraviolet observations useful for studying parts of the sky that may not be apparent in other kinds of light.
There is a basic obstacle to building a full map from the ground: the ozone layer blocks ultraviolet light. That means it can only be measured from space. A complete ultraviolet map had not existed before this project.
NASA’s GALEX mission had already surveyed about two-thirds of the sky. But it left out bright star-forming regions, so the available coverage did not provide a complete picture. Combining observations from multiple missions offered a way to bring together data that had not previously formed one full map.
How Claude Science assembled the map
Johns Hopkins astrophysicist Brice Ménard describes the process on the Anthropic website. Claude Science coordinated AI agents that downloaded data from multiple space missions, calibrated the observations and merged them together.
These steps turn separate mission data into a shared resource. Calibration prepares measurements to be combined, while merging brings observations into one map. The project therefore relied on a sequence of data tasks, rather than on a single model simply generating a picture of the sky.
Some areas were still missing. To fill those gaps, the team used inpainting: a technique in which a model learns from existing data and reconstructs absent sections. The missing areas were predictions, rather than direct measurements. In tests, those predictions averaged about ten percent deviation from actual measurements.
A teaching resource and a possible research pattern
The map is intended as teaching material. Its creation also offers an example of AI agents handling the practical work involved in assembling a scientific dataset: collecting material from different missions, calibrating it, joining it and estimating gaps where observations were absent.
Ménard suspects that many scientists have postponed similar projects. The implication is that work that once seemed too tedious to complete may become more achievable when AI can coordinate these steps. That does not establish that every such project is suitable for AI, or that reconstructed data is interchangeable with observed data. In this case, the map used inpainting for missing regions, and tests compared its predictions with actual measurements.
The distinction matters for anyone using the map. It combines observations from space missions with modeled estimates in areas where data was missing. Knowing how those parts were produced helps readers understand what the map can show and where its coverage depends on reconstruction.
For researchers and students, the project provides both a map and an account of how it was assembled. Its broader lesson, as Ménard presents it, is that AI may make certain labor-intensive data projects possible to pursue. The work still depends on space-based ultraviolet measurements and on checking how well predictions match actual observations.