AI Wire

Short briefings on models, tools, agents, and the rules that govern them.


English edition

Models

Astrophysicist Uses Claude to Map Ultraviolet Sky

Johns Hopkins and Anthropic research combines archival satellite surveys with AI models to fill an unobserved third of the ultraviolet sky.

Research from Johns Hopkins and Anthropic merges archival GALEX data with statistical models to cover blind spots along the Milky Way.

A research report published by Anthropic introduces the first continuous all-sky map rendered in ultraviolet light. The project was led by Brice Ménard, an astrophysics professor at Johns Hopkins University and a researcher at Anthropic. To complete the panoramic view, the team used the Claude Science platform to orchestrate pipeline tasks and mathematically reconstruct roughly 33% of the celestial sphere that space telescopes never directly observed.

Gaps Left by Space Telescopes

Earth’s ozone layer blocks almost all incoming ultraviolet radiation, forcing astronomers to rely on orbital observatories. NASA’s Galaxy Evolution Explorer (GALEX), which operated from 2003 to 2013, collected roughly 38,000 exposures covering nearly two-thirds of the sky. The instrument deliberately avoided pointing at the dense core of the Milky Way and bright stars to prevent sensor damage.

To assemble archival catalogs from GALEX, Swift-UVOT, FIMS/SPEAR, and TD-1, Ménard deployed software agents within Claude Science. The tools handled data downloads, coordinate reprojections, and stellar calibration across more than 100 million stars cataloged by the European Space Agency’s Gaia mission. Engineering routines that typically require weeks of manual coding were wrapped up in a few days.

How Claude Estimated the Ultraviolet Sky

The remaining third of the map did not come from newly launched hardware. Instead, the model estimated emissions in the far-ultraviolet (154 nanometers) and near-ultraviolet (232 nanometers) bands by analyzing correlations with optical, infrared, and radio sky surveys. Cross-validation runs showed an average relative prediction error between 10% and 14%.

The release draws an explicit line between physical observations and synthetic estimates. In the data files, each pixel is labeled as measured or predicted, accompanied by statistical uncertainty intervals. The dataset is distributed under a Creative Commons license (CC BY 4.0), while the data pipeline scripts are open source under the MIT license.

Practical Limits and Human Supervision

Despite agentic speedups, the project required direct human verification. Ménard had to manually inspect the stitched mosaic to catch visual artifacts that automated scripts failed to flag. Additionally, the project appeared as an institutional technical report on Anthropic’s blog rather than a peer-reviewed academic journal article, reflecting the author’s dual academic and corporate affiliation.

The team positions the map as an educational guide and reference model rather than a replacement for physical sky surveys. Algorithmic estimates provide plausible baselines, but they cannot verify physical discoveries that demand empirical proof. Direct empirical coverage across the entire ultraviolet spectrum will likely wait until roughly 2030, when NASA plans to launch its UVEX space observatory.

Sources

Sources

  1. anthropic.com
  2. x.com
  3. alphasignal.ai
  4. superpowerdaily.com

Markers

  • astronomy
  • claude-science
  • anthropic
  • galex
  • artificial-intelligence
  • uvex

Corrections and editorial policy