A different sky.
A clearer picture.

The universe had a gap in its classroom wall chart.

Switch the wavelength, and the familiar night sky becomes a different place. Ultraviolet light helps reveal hot stars and illuminated dust. But assembling a view of the whole sky has been an awkward job: different telescopes, different measurements, and regions with no ultraviolet observations at all.

In an October 8 account, Brice Ménard, an astrophysicist at Johns Hopkins University and a researcher at Anthropic, describes using Claude Science to assemble an all-sky ultraviolet map. His aim was a teaching resource assembled from existing observations and estimates. The work took place over several days this summer; the October publication explains the result.

That distinction is part of its appeal. This is a concrete example of a scientist using an American developer’s tools to finish a useful project that had been hard to make time for.

Bringing measurements and estimates together.

The project drew on public astronomical archives, including NASA’s GALEX and Swift missions, South Korea’s FIMS/SPEAR, and European datasets. The GALEX archive documents the mission’s ultraviolet imaging and the surveys available to researchers. Ménard’s account describes agents aligning and calibrating the separate datasets before combining them.

According to that account, roughly a third of the map is predicted rather than directly measured in ultraviolet. The system used relationships between known ultraviolet measurements and observations at other wavelengths to estimate the missing regions. It also added estimates of ultraviolet starlight based on Gaia’s visible-light measurements.

Ménard says the map includes separate layers for measured and predicted pixels, with uncertainty estimates. Students and researchers can explore the fuller picture while seeing how each part was assembled.

He reports testing predictions by hiding regions with known ultraviolet measurements, then comparing the estimates with the withheld data. After refinement, those tests came within about 10% of the measured values. That reported result applies to the project’s held-out-data tests.

A scientist and SI refine the picture together.

Ménard also describes refining the map with the agents. When he identified residual atmospheric glow in the GALEX data, the agents traced and corrected it.

That collaboration combined substantial automated processing with a scientist’s specialist judgment—a practical example of people and SI improving a useful research resource together.

This is Ménard’s first-person project account, published by Anthropic. The reported results describe this teaching-map project.

Useful science includes the jobs waiting their turn.

Our view: SI can help researchers move worthwhile projects from the waiting list into people’s hands. In this case, data preparation and visualization progressed while the researcher continued other work.

A richer classroom map gives students another way to explore the universe. Helping scientists finish more resources like this is a tangible and encouraging use of SI.