Michael JohnsonAbout a third of the sky has never been observed in ultraviolet light. On October 8, 2026, Anthropic published astrophysicist Brice Ménard’s account of how he used Claude Science to combine archival space observations and estimate the missing areas. The map distinguishes measured regions from model predictions, which also include uncertainty estimates.
The project shows how AI agents can help with the labor-intensive processing of scientific archives: finding datasets, bringing observations onto a common scale, and running repetitive computations. To test the predictions, Ménard hid portions of regions with known ultraviolet measurements from the model, then compared the reconstructed values with the actual data.
How observations from different telescopes were combined
Ultraviolet light is absorbed by atmospheric ozone, so these observations are made from space. The project’s main dataset came from NASA’s GALEX mission, which observed from 2003 to 2012. The MAST archival survey reports that GALEX observed 77% of the sky in at least one band; the mission deliberately skipped some areas containing very bright stars because of the risk to its detectors. The project also drew on additional data from Swift and South Korea’s FIMS/SPEAR mission.
Observations from different instruments could not simply be added together: they had to be cross-calibrated, brought to the same resolution, and aligned to a shared coordinate system. According to Ménard’s account, Claude coordinated agents that searched open astronomical surveys, processed images, and combined the results. In the project’s workflow, components produced by deterministic procedures are distinguished from the model’s statistical estimates.
Gaps were estimated using other wavelength bands
To fill the blank regions, Claude used inpainting—a method for estimating missing data from patterns in existing images. Data in visible, infrared, and radio wavelengths were available for areas without ultraviolet observations. In regions where ultraviolet measurements did exist, the model estimated the relationship between them and observations in other bands, then applied those relationships to areas not covered in ultraviolet.
Ménard tested the predictions on known regions: he hid some real ultraviolet data and asked the system to reconstruct it. After several iterations, the estimates differed from the hidden measurements by about 10% in this test. That result describes testing on known regions; Anthropic also says the completed map has separate layers marked “measured” and “predicted,” along with uncertainty estimates.
The map also includes estimates of ultraviolet light from more than 100 million individual stars. These were derived from visible-light measurements by the European Space Agency’s Gaia mission. The map thus combines archival observations and calculated layers while retaining information about their origin.
Human review caught an error after two rounds of review
While inspecting the images, Ménard noticed faint circular traces from individual GALEX exposures. They arose from residual ultraviolet glow in Earth’s atmosphere: when the background is not fully removed, the circular observation fields differ slightly in brightness. According to Ménard’s account, two rounds of review by other agents did not catch the defect. After he pointed it out, the agents corrected the background in all 38,000 observations.
This episode shows that agents sped up repetitive operations, but visual inspection by a specialist remained part of the workflow. In the final setup, it is useful to preserve intermediate layers, distinguish predicted regions, and check the results against observations and through visual analysis.
Anthropic presents the map as an educational resource that clearly shows structures of the Milky Way in ultraviolet light. The project’s practical takeaway is that agent-based processing can help bring labor-intensive work with open archives to completion, provided measurements, model estimates, and expert-review findings remain distinguishable.