Anthropic Claude Science Generates First Complete Ultraviolet Sky Map

TL;DR

Anthropic’s Claude Science AI orchestrated the collection, calibration, and inpainting of multiple ultraviolet (UV) sky surveys to produce the first complete UV map of the entire sky, filling gaps that have existed for decades and providing a new educational tool for astrophysics.

Why a Complete UV Map Was Missing

UV observations require space‑based telescopes because Earth’s ozone absorbs UV light. Over the past 50 years, missions such as NASA’s GALEX (2003‑2013) imaged roughly two‑thirds of the sky but deliberately avoided bright‑star regions, especially along the Milky Way plane, to protect detectors. Additional UV data from NASA’s Swift, South Korea’s FIMS/SPEAR, TD‑1, and others still left substantial unobserved patches—about one‑third of the celestial sphere. Traditional statistical gap‑filling methods demand weeks of meticulous pixel‑level calibration, a labor‑intensive effort that most researchers set aside for higher‑priority work.

How Claude Science Built the Map

Data Gathering and Normalization

  • Claude deployed a fleet of AI agents to locate and download every publicly available UV survey, including ~38,000 GALEX images and datasets from Swift, FIMS/SPEAR, and TD‑1.
  • Agents performed per‑image calibration to remove star‑induced glare and atmospheric background, ensuring consistency across observations taken at different times and with different instruments.

Cross‑Calibration and Merging

  • The agents resampled each survey to a common resolution and projected all data onto a unified celestial coordinate grid.
  • Cross‑calibration aligned the photometric scales of the disparate instruments, enabling a seamless merge into a single UV mosaic.

Inpainting the Unobserved Third

  • Claude applied a machine‑learning inpainting technique, training on the two‑thirds of the sky that already had UV measurements.
  • The model learned statistical relationships between UV intensity and complementary wavelengths (visible, infrared, radio) and used these correlations to predict UV values for the missing regions.
  • Validation involved hiding known UV data, inpainting it, and measuring error; the model achieved ~10 % deviation from real measurements, a difference imperceptible to the human eye.

Adding Stellar Contributions

  • Using Gaia DR3 visible‑light measurements, Claude estimated UV emission from more than 100 million individual stars and incorporated these point sources into the final map.

Iterative Refinement

  • Early versions displayed residual circular artifacts—remnants of GALEX pointings with incomplete atmospheric‑glow subtraction.
  • After identifying the issue, Claude re‑estimated the background for all 38,000 GALEX observations, eliminating the disc‑shaped artifacts.
  • Over a dozen successive versions were produced, each refined through human‑AI interaction cycles.

The Final Product

  • The completed UV sky map combines far‑UV (154 nm) and near‑UV (232 nm) data, covering the entire celestial sphere.
  • Visible structures include dust clouds illuminated by young stars, giant dust rings from supernova‑driven bubbles, faint interstellar filaments, and the Large and Small Magellanic Clouds.
  • The map is publicly accessible at https://menard.pha.jhu.edu/uvmap/ and includes per‑pixel labels indicating whether the value is measured or predicted, along with uncertainty estimates.

Implications for Education and Research

  • The map provides a missing wavelength layer for teaching multi‑spectral astronomy, allowing students to directly compare UV, visible, infrared, and radio views of the Milky Way.
  • Researchers gain a baseline UV dataset for studies of dust scattering, star‑formation regions, and galactic‑scale UV background radiation.
  • The workflow demonstrates how AI‑orchestrated pipelines can tackle low‑priority, data‑intensive projects that would otherwise remain unfinished.

Acknowledgments

  • Data sources: GALEX, Swift‑UVOT, FIMS/SPEAR, TD‑1, Planck, and ESA Gaia DR3.
  • Many Claude Science agents performed the bulk of the data processing; the author guided the overall strategy.

All statements are based on the Anthropic research post dated 2026‑10‑08.

Sources