Claude Science UV sky map fills observation gaps with labeled predictions

Astrophysicist Brice Ménard says he worked with Claude Science to assemble an ultraviolet map covering the whole sky. About a third of the image is predicted rather than directly observed in ultraviolet light, with additional layers identifying measured and predicted pixels and providing uncertainty estimates.

That distinction is central to the account published by Anthropic. Ménard, a Johns Hopkins University astrophysicist and Anthropic researcher, describes the map as an educational resource. The work combines existing observations with statistical gap filling; it does not turn unobserved regions into new telescope measurements.

Combining surveys required calibration before prediction

Ménard explains that ultraviolet observations require space-based instruments because atmospheric ozone absorbs UV light. NASA's GALEX mission supplied the largest dataset, imaging around two-thirds of the sky in approximately 38,000 observations between 2003 and 2013. It avoided some bright-star regions to protect its detectors, leaving gaps that other surveys did not completely fill.

His instructions to Claude Science were to collect available datasets, put them on a common scale, merge them and estimate the remaining regions. He says agents searched for public surveys and downloaded observations taken at different times and under varying conditions.

The agents then had to make each survey internally consistent, including dealing with glare around bright stars. Combining instruments required cross-calibration, a common resolution and a shared coordinate system. Those steps matter before interpreting differences in a map as features of the sky rather than differences between surveys.

Ménard describes supplying high-level instructions while agents performed the processing. The account is a case study of his collaboration with Claude, not an independent comparison against another pipeline or a claim that calibration can proceed without scientific guidance.

Missing ultraviolet regions were estimated using other wavelengths

To fill the gaps, Ménard asked Claude Science to use inpainting informed by observations in visible, infrared and radio wavelengths. The method learned relationships between those observations and UV brightness in the measured part of the sky, then used them to predict the unobserved third.

He reports testing the predictions by hiding portions of regions where UV measurements already existed. After several refinement rounds, the model estimated hidden values to within about 10% of the real UV measurements. That is the author's reported validation result for those tests, not a guaranteed error bound for every unobserved region.

The account also describes adding estimated UV light from more than 100 million stars using visible-light measurements from ESA's Gaia satellite. Both that contribution and the diffuse inpainted background involve inference. A complete-looking image therefore needs its measurement-status and uncertainty layers to remain available to users.

Human inspection caught an error that agent reviews missed

Ménard recounts finding faint circular patterns in a dim field. They were footprints of individual GALEX observations, caused by residual atmospheric glow. Claude had identified the issue as a known risk, but two rounds of review by other agents had not caught it in the map.

After Ménard pointed out the circles, the agents traced the problem and corrected the background across the observations. He says the collaboration produced more than a dozen versions over several days, with periods of discussion followed by hours of computation.

That episode supplies a concrete limit on the account's automation: agreement between reviewing agents did not establish that the output was free of processing artifacts. The scientist's inspection changed the result.

The finished map can help students compare UV structures with views at other wavelengths. Its usefulness rests on keeping estimated regions distinct from measurements and preserving uncertainty, while documenting the calibration decisions behind the image. The case illustrates a productive scientific-assistance workflow without establishing that every inferred structure is a directly observed discovery.