How Claude Helped an Astronomer Build a Complete Ultraviolet Map of the Sky

Anthropic recently shared a science story with a cosmic scale: Brice Ménard, an astrophysicist at Johns Hopkins University and a researcher at Anthropic, used the company's Claude Science workspace to build what the post describes as the first complete ultraviolet map of the entire sky.

Here is the striking part: roughly one-third of that sky had never been observed by an ultraviolet telescope. Those regions were filled in with model predictions. Each pixel is labelled to distinguish observations from predictions, with an accompanying uncertainty estimate.

According to Anthropic's account, work that would normally take weeks was completed in days. While Claude handled much of the processing, Ménard continued working on other projects. This is AI for science, or AI4S, in a very tangible form.

Why the Ultraviolet Sky Still Had Gaps

Ménard often begins his astrophysics classes by showing students the same sky at different wavelengths. In visible light, stars dominate. In infrared, dust grains glow along the band of the Milky Way. Switch to radio wavelengths, and stars and dust recede while hydrogen gas, the raw material for new stars, comes into view. In X-rays, the sky reveals energetic phenomena such as stellar explosions and matter spiralling into black holes.

The sky in visible light, infrared and radio wavelengths.

Ultraviolet light has its own stories to tell: dust illuminated by starlight, clouds surrounding young stars, and rings left behind by stellar explosions.

But when Ménard reaches this part of the lesson, he says he always has to offer an apology. The ultraviolet maps have holes.

The ozone layer absorbs ultraviolet light, making space-based observations necessary. Even in space, however, different telescopes have surveyed different patches of sky rather than producing a single, complete picture.

The largest dataset comes from NASA's Galaxy Evolution Explorer, GALEX. Between 2003 and 2013, it made approximately 38,000 observations, covering about two-thirds of the sky. To protect its detectors from intense light, it deliberately avoided regions crowded with bright stars, including the entire plane of the Milky Way.

NASA's Swift and South Korea's FIMS/SPEAR filled in some additional regions. Combined, their observations still left substantial gaps.

Stitching those datasets together was technically possible, but painstaking. It required pixel-by-pixel calibration, and astronomers usually had more urgent research to pursue.

In other words, it was exactly the kind of slow, low-priority work that keeps getting pushed to the bottom of the list.

Asking Claude to Fill in the Sky 🔭

That summer, Ménard decided to tackle the problem with Claude Science. The brief was simple to describe, if not to execute:

Find all the available ultraviolet data, bring it onto a common scale, combine it into one map, and fill in every region that ultraviolet telescopes have never observed.

Claude's workflow unfolded in four main stages:

1. Find and download public ultraviolet survey data. GALEX alone contributed tens of thousands of images, collected over many years under varying conditions.
2. Clean the datasets. Make each dataset internally consistent and address artefacts around bright stars, including glare.
3. Align and cross-calibrate the observations. Bring data from multiple instruments onto a common coordinate system, resolution and measurement scale.
4. Predict the missing regions. For the roughly one-third of the map without observations, Ménard asked Claude to use inpainting, a technique for reconstructing missing portions of an image from learned patterns.

After several rounds of refinement, the account reports that the reconstructed values differed from real measurements by around 10% in comparisons. Visually, the difference was difficult to spot. That does not turn a prediction into an observation, which is why the map keeps the two clearly distinguished.

Claude then used visible-light measurements from the European Space Agency's Gaia satellite to estimate the ultraviolet brightness of more than 100 million stars. These were layered onto the reconstructed background, producing the first version of the complete all-sky ultraviolet map.

From a patchwork of missing regions to a complete map, in four main stages.

The pipeline diagram brings the process together: cylinders represent public datasets, blue boxes show rule-based steps executed by Claude, purple boxes represent statistically learned components, and green boxes show the merging process and final output.

Crucially, the merging stage followed a hierarchy: GALEX observations took priority, while model predictions came last. Where measured data existed, predictions did not override it.

A Human Still Had to Spot the Problem

Claude did not get everything right.

One evening, while inspecting the processed maps, Ménard noticed something odd in one of the darkest regions: faint circles, each slightly brighter or darker than its neighbours.

These were the footprints of individual GALEX exposures. Each observation covered a circular patch of sky and contained a slightly uneven contribution from ultraviolet airglow in Earth's atmosphere. When that glow was not fully subtracted, the boundaries between exposures became visible.

Claude had listed this issue among the project's known problems at the outset. Yet the map still passed two rounds of review by other agents. It took Ménard's own visual inspection to catch the residual artefacts.

His instruction was brief:

“I can see individual discs, the imprints of separate observations. Can you fix that?”

The agents traced the problem back to residual airglow, and Claude recalibrated all 38,000 observations. A few hours later, the circles were gone.

Before and after correction: residual observation footprints on the left, corrected results on the right.

More Than a Dozen Major Iterations

The project ran over several days and went through more than a dozen major versions. The working rhythm was straightforward:

• Ménard and Claude Science exchanged a few messages to decide what to do next.
• Claude worked through several hours of computation while Ménard returned to other projects.
• They reviewed the results and repeated the cycle.

The final map offers a remarkable view of the universe: glowing dust clouds around bright young stars; vast dust rings left by bubbles driven outward by exploding massive stars; and faint filaments far from the Milky Way's plane, illuminated by the combined starlight of the galaxy.

Below the galactic plane sit the Large and Small Magellanic Clouds, two neighbouring galaxies rich in young stars.

In his acknowledgements, Ménard thanked the people behind the space missions who made their data public, as well as the Claude Science agents that carried out the patient, painstaking work.

As for his own role, he put it simply:

“My contribution was limited to guiding them.”

The Bigger Picture

The original account notes that Claude Science opened on June 30 and was still in beta at the time of writing.

The more interesting takeaway is the division of labour. AI took on the repetitive, data-heavy processing. A scientist set the direction, checked the results and caught an error that automated reviews missed.

A complete map does not mean a completely observed sky. Its value lies in bringing measurements and explicitly labelled predictions together, while keeping their uncertainties visible.

For AI4S, that is a promising picture: not science without scientists, but scientists with more room to do science. ✨