Claude models the ultraviolet sky’s missing third

Astronomers have all-sky maps in radio waves, infrared, visible light, X-rays and gamma rays. Ultraviolet coverage remained patchy because Earth’s ozone layer absorbs UV radiation, leaving observations to space telescopes that had skipped large sections of the sky.

Johns Hopkins astrophysicist Brice Ménard used Claude Science, Anthropic’s workbench for planning and executing multistep research tasks, to assemble and run a reconstruction pipeline. The resulting first full-sky UV map combines far-ultraviolet light at 154 nanometers with near-ultraviolet light at 232 nanometers.

Every pixel carries its provenance

The map’s completeness refers to spatial coverage. Scientific analyses can use the provenance and uncertainty fields to filter, weight or separately evaluate modeled pixels. The finished view traces dust clouds around young stars, large rings formed by stellar explosions and faint filaments illuminated by combined galactic starlight.

Why UV surveys stopped short

NASA’s GALEX mission supplied the largest existing UV dataset, capturing roughly 38,000 observations from 2003 to 2013 and covering about two-thirds of the sky. Mission planners avoided locations with very bright stars, especially along the Milky Way’s plane, because intense light could damage the spacecraft’s detectors.

NASA’s Swift observatory and South Korea’s FIMS/SPEAR mission covered additional regions, yet substantial gaps remained. Combining those archives requires pixel-level calibration, artifact removal, coordinate conversion and repeated statistical analysis. Such projects can consume weeks of specialist time, which often leaves them behind research with firmer deadlines.

Prompts became a processing pipeline

Ménard supplied high-level instructions, inspected intermediate products and requested corrections. Claude coordinated parallel sub-agents and long-running computations across the following stages:

1. Archive ingestion: Locate public UV surveys and download tens of thousands of source images.
2. Artifact removal: suppress glare around bright stars so nearby faint emission remains measurable.
3. Instrument calibration: reconcile brightness measurements from different telescopes.
4. Coordinate alignment: reproject every exposure onto a shared sky grid and merge overlapping observations.
5. Statistical reconstruction: estimate missing UV emission from visible, infrared and radio measurements of the same regions.
6. Stellar modeling: add UV estimates for more than 100 million stars using visible-light data from ESA’s Gaia mission.

The reconstruction stage used observed regions to model how ultraviolet brightness relates to emission at other wavelengths. It then applied those relationships to unobserved pixels and generated a confidence estimate for each prediction, a process commonly called inpainting.

Ménard tested the method by masking regions with known UV measurements and asking the pipeline to reconstruct them without access to the hidden values. After several rounds of refinement, the reported predictions differed from the withheld measurements by about 10%.

Faint circles escaped two reviews

Intermediate maps contained subtle circles in the dimmest fields, with each circle appearing slightly brighter or darker than its surroundings. The patterns traced individual GALEX exposures whose uneven atmospheric UV glow had survived the initial corrections.

Claude had identified atmospheric glow as a known risk early in the project, yet two rounds of agent review failed to detect its imprint on the combined map. Ménard spotted the circles during visual inspection and prompted the system to correct all 38,000 observations, which took a few hours.

Global visualization provided a quality check that local processing and automated review had missed. Similar pipelines can test for tile boundaries, background offsets and repeated detector footprints before accepting a merged data product.

Patterns developers can reuse

• Separate deterministic and statistical stages. Downloads, calibration and reprojection can produce reproducible outputs before learned components estimate missing values.
• Parallelize independent regions. Image tiles and sky sectors can be processed concurrently, reducing elapsed time without changing the underlying method.
• Preserve provenance. Downstream users need to know which values came from instruments and which came from inference.
• Attach uncertainty to predictions. A filled map remains useful when applications can account for confidence at pixel level.
• Validate against hidden ground truth. Masking measured regions creates controlled tests for reconstruction accuracy.
• Inspect aggregate artifacts. Some calibration failures become visible only after thousands of inputs are combined.

The workflow produced more than a dozen map versions over several days. Ménard planned each iteration in short exchanges, then allowed the system to run computations while he worked elsewhere. For data teams, the project offers a concrete example of domain expertise directing agents through a dependency graph of ingestion, transformation, inference and review.

Provenance defines the map’s limits

The missing regions form a systematically different sample because GALEX deliberately avoided bright stars and much of the galactic plane. Validation on masked portions of the observed sky may therefore understate errors in regions whose structure, brightness and source density differ from the training data.

Research use requires more detail than the headline 10% result provides, including the error metric, test-mask design, regional performance, uncertainty calibration and behavior near bright sources. Reproducibility also depends on recording code, model versions, prompts, calibration parameters, archive versions and compute requirements.

The project demonstrates how supervised agents can make a deferred data-integration project feasible within days. Its measured-versus-modeled labels, held-out tests and visible failure history also show the controls required when statistical reconstruction becomes part of a scientific dataset.