Brice Ménard, an astrophysicist at Johns Hopkins University and a researcher at Anthropic, described a map spanning the ultraviolet sky on October 8. Dated September 16, it combines space-telescope observations and marks which parts were measured, translated from another ultraviolet band, or estimated. AI agents helped find, clean and calibrate the data; a conventional statistical model supplied the remaining estimates.
NASA’s GALEX mission supplied most of the observations through about 38,000 snapshots taken from 2003 to 2013. It avoided very bright stars and parts of the Milky Way to protect its detectors. By itself, GALEX covered 64% of the sky in far-ultraviolet light and 76% in near-ultraviolet light; after 2009 it observed only in the near-ultraviolet band. Swift-UVOT, Korea’s FIMS/SPEAR instrument and the older TD-1 satellite contribute additional measurements.
Ménard’s team put the surveys on a common brightness scale and kept measured values intact. Coarser readings, such as FIMS/SPEAR’s roughly one-degree far-ultraviolet map, take priority over a model estimate. An ensemble of small decision trees predicts the diffuse ultraviolet glow where finer data remain missing, using maps of dust, hydrogen emission and Gaia starlight. The model contributes to 28% of the far-ultraviolet sky and 27% of the near-ultraviolet sky. Those are the shares with model-estimated detail, not areas without any ultraviolet measurements: much of the far-ultraviolet sky without GALEX images has a coarse FIMS/SPEAR reading, while its finer structure remains predicted. The map also adds ultraviolet flux estimates for about 119 million Gaia stars.
Those labels matter for resolution as well as provenance. The 1.7-arcminute grid is the sampling of the map, but predicted diffuse regions have an effective resolution of about 0.5 to 1 degree. Ménard warns that estimated areas cannot support discoveries of new sources and do not provide an independent measure of interstellar dust.
To test the estimates, Ménard hid observed GALEX patches and compared predictions with the withheld measurements. His map page reports typical prediction errors of 12% to 14% on these patches, growing farther from measured data. For context, it gives about 5% as the uncertainty in GALEX-imaged areas; that is a measurement uncertainty, not a second withheld-patch prediction test. The page says its validation has not been independently reviewed by people, and the technical paper remains in preparation.
The workflow also needed a human to catch a flaw. Ménard says two rounds of agent review missed faint circular artifacts left by individual GALEX observations. After he spotted them, Claude’s agents traced them to residual atmospheric glow and corrected all 38,000 observations.
Sources
- Brice Ménard: map methods, measured-versus-predicted layers and limits
- Anthropic: Ménard’s October 8 account of the AI-assisted workflow
- NASA/JPL: the GALEX ultraviolet image of M94 used above
- JPL image-use policy and credit requirements
Image: NASA/JPL-Caltech. The M94 image is an example of GALEX ultraviolet observations, not the new full-sky map.