Ménard's Claude Science agent team filled a third of the unobserved sky to produce the first complete UV all-sky map

The work produced the first complete map of the sky in ultraviolet light, covering the whole sky, with about one third of it, including much of the galactic plane, predicted rather than measured. Previously the only UV map available for teaching was full of holes: GALEX imaged about two-thirds of the sky in some 38,000 observations between 2003 and 2013 but deliberately skipped bright-star regions, including the plane of the Milky Way, to avoid damaging its detectors, and combining Swift and FIMS/SPEAR data still left missing pieces. The map carries additional layers labeling each pixel as measured or predicted and providing uncertainty estimates, so predicted and measured regions can be distinguished and traced.

Missing sky was estimated by inpainting combined with multi-wavelength information: the model learned how UV brightness relates to visible, infrared and radio observations over the two-thirds of the sky already mapped in UV, then applied that relationship to the third with no UV data, estimating each point and its confidence. It pairs machine-learning inpainting with astronomical multi-wavelength relations, replacing a statistical completion process that otherwise requires weeks of pixel-level calibration and repeated analysis. The author tested this by deliberately hiding parts of regions with existing UV data; after several rounds of refinement the model estimated the hidden data to within about 10% of the real UV measurements, a difference described as almost imperceptible to the human eye.

On top of the diffuse inpainted background, the map adds estimates of UV light from more than 100 million individual stars inferred from visible-light measurements by ESA's Gaia satellite. Layering diffuse UV background with per-star contributions lets the final map show dust clouds glowing around young stars, giant dust rings and loops tracing bubbles from massive-star explosions, faint dust filaments far from the galactic plane, and the Large and Small Magellanic Clouds. The stellar layer is inferred from public Gaia measurements as part of the map-building pipeline; the text does not report a separate accuracy test for the per-star estimates.

The pipeline was executed by multiple AI agents orchestrated through Claude Science: searching for and downloading public UV survey data, removing glare around bright stars to make each survey internally consistent, cross-calibrating instruments and unifying resolution and coordinate system, then inpainting and layering. The author notes that such pixel-level calibration work is usually set aside as lower priority, whereas here more than a dozen successively refined map versions were produced over several days with his contribution limited to high-level guidance. Concrete process evidence is given: after the agents corrected leftover atmospheric glow across all 38,000 GALEX observations, the circular observation footprints the author could see by eye disappeared; the author also records that the map passed two rounds of agent review before he spotted the problem.

The map is intended for teaching and structural display, and for astronomical users who need complete UV coverage as a reference; it applies where public survey data are available and where a learnable multi-wavelength relationship exists between missing and observed regions. Methodologically, the inpainting relies on auxiliary visible, infrared and radio observations, so in regions where those are also missing or the relationship breaks down, prediction reliability must be judged separately. The author also notes that this pixel-level calibration work was previously set aside because it took weeks, and that with agent assistance it advanced within days, offering a replicable way to organize backlogged map and resource projects in other fields.

Readers should note that about one third of the sky's UV brightness is model-estimated rather than measured, and its credibility depends on whether the multi-wavelength relationship holds in regions with no UV data, a question the map's per-pixel uncertainty layer is meant to carry. The held-out test reports about 10% difference, but it was constructed inside regions that already have UV data, and whether conditions there match truly unobserved regions is not developed in the text. Final map quality also depends on how completely systematic effects such as leftover atmospheric glow are removed; the circular-footprint problem the author records suggests such effects can remain in visually detectable form.