Anthropic's Claude builds first complete UV sky map
Using Anthropic's Claude Science, an astrophysicist has mapped the missing third of the ultraviolet sky, demonstrating how AI agents can automate complex scientific data integration.
Johns Hopkins astrophysicist Brice Ménard utilized Claude Science, Anthropic's agentic research workbench, to construct the first complete map of the ultraviolet sky. Historically, roughly one-third of the sky—including much of the Milky Way's galactic plane—remained unobserved in UV wavelengths because NASA's GALEX space telescope bypassed bright-star regions to protect its sensitive detectors. By combining far-ultraviolet light at 154 nanometers and near-ultraviolet light at 232 nanometers, the new map successfully fills this massive observational gap.
To build the map, Claude coordinated multiple parallel sub-agents to execute a complex multi-stage pipeline. The AI ingested, calibrated, and merged massive datasets from GALEX, Swift, FIMS/SPEAR, TD-1, Planck, and Gaia. This process involved downloading tens of thousands of images, aligning coordinates onto a shared sky grid, and reconciling brightness measurements across different instruments. To fill the unobserved regions, the system used statistical inpainting, mapping correlations between known UV emissions and visible, infrared, and radio wavelengths. It also inferred UV estimates for more than 100 million stars using visible-light data from the European Space Agency's Gaia mission.
The reconstruction proved highly accurate. In blind validation tests where researchers masked known UV data, Claude's predictions came within approximately 10% of the actual measurements. However, the project still required human oversight. During visual inspection, Ménard noticed faint circular artifacts caused by uneven atmospheric glow in the original GALEX exposures, which automated reviews had missed. After Ménard flagged the issue, Claude corrected all 38,000 GALEX observations within a few hours.
The entire workflow generated more than a dozen map iterations over just a few days, demonstrating how supervised AI agents can make labor-intensive data-integration projects feasible in a fraction of the usual time. For practitioners, the resulting dataset provides crucial provenance and uncertainty labels for every pixel, allowing researchers to filter or weight modeled data in future astronomical analyses.
This is our own summary of reporting by AlphaSignal
