Claude Science Builds First Complete UV Map of the Sky

Anthropic says its Claude Science system has produced something astronomers did not have before: a map of the entire sky in ultraviolet light. Brice Ménard, an astrophysicist at Johns Hopkins, describes the project on Anthropic's website. The company presents it as an example of AI handling the slow, repetitive data work that researchers often never find the time to finish.

The result is interesting in its own right. It is also a useful case study in where AI agents fit into scientific work today.

Why the ultraviolet sky was never fully mapped

Ultraviolet observations show material that other wavelengths can miss. In UV, astronomers can see dust lit up by starlight. Examples include clouds surrounding young stars and the rings that stellar explosions leave behind.

The difficulty is getting the data. Earth's ozone layer blocks UV light, so ground-based telescopes can't capture it. Every measurement has to come from instruments in space, which sharply limits how much data exists.

NASA's GALEX mission did the most comprehensive work so far, covering about two-thirds of the sky. It left out bright star-forming regions, so large gaps remained in the UV picture of the sky. Until this project, nobody had a complete map.

How Claude Science put the map together

Claude Science coordinates a group of AI agents, and the work was split into three main jobs:

1. Collecting the data. The agents downloaded observations from several space missions.
2. Calibrating it. Data from different instruments has to be adjusted before it can be compared, and the agents handled that step.
3. Merging it. The calibrated datasets were combined into one map.

That still left holes where no usable observations existed. To fill them, the agents used inpainting. In this technique, a model learns from the data that is available and then reconstructs what the missing regions most likely contain.

According to the account, the team tested how well this worked. On average, the predicted values differed from real measurements by about ten percent. That figure matters. The filled-in sections are informed estimates, not direct observations, and anyone using the map should keep the difference in mind.

A map for teaching

The stated goal is modest. The map is intended as teaching material, not as a replacement for dedicated observations. That framing fits the error margin: a ten percent average deviation is reasonable for showing students the structure of the UV sky, while precision research would still need real measurements.

Ménard's broader point may matter more than the map. He suspects that many scientists have projects like this one sitting on the shelf. These jobs are worth doing but require so much downloading, cleaning and stitching of data that nobody gets to them. AI agents, in his view, could now make some of that work feasible.

The pattern behind the project

This fits a wider move toward using language models as research assistants. Open-source efforts such as BootLoops, a harness that turns Claude into a lab tool, point in the same direction. So do projects built on large public science archives, like the NASA-IBM lunar foundation model. Each one tries to turn a large, messy pile of scientific data into something researchers can work with.

Our Take

The UV map is a modest result, and that is what makes it convincing. The project does not claim new physics. It shows agents taking on the data plumbing that slows down real research: finding files, reconciling instruments and joining datasets. For readers who follow AI agents, this suggests that their most practical near-term value in science lies in clearing backlogs, not in replacing scientists.

The ten percent deviation is also a reminder that reconstructed data carries uncertainty. As more AI-built datasets appear, it will be worth watching whether teams report their error margins as clearly as this one did. A related question is how they will mark reconstructed regions, so that downstream users can tell estimates from measurements.

The next test is whether outside researchers pick up Ménard's suggestion and use similar agent workflows on their own delayed projects. Independent results of that kind would show whether this is a repeatable method or a well-chosen showcase.