AI Fills Cosmic Gaps: First Complete Ultraviolet Sky Map Unveiled

On October 8, Anthropic released something astronomers have wanted for decades but never quite managed to assemble: a complete map of the ultraviolet sky. The catch is that roughly one-third of it was never actually seen by any telescope. Instead, an artificial intelligence model inferred those missing pieces by learning patterns from the wavelengths that were observed.

The project emerged from a practical frustration. Brice Ménard, an astrophysicist at Johns Hopkins University who also works with Anthropic, wanted to show his students a gap-free ultraviolet map of the cosmos. The problem was that no such thing existed, and the work required to create one—sifting through decades of satellite data, standardizing measurements from different instruments, and stitching everything together—had always seemed too laborious to justify. This summer, he decided to hand the task to Claude Science, Anthropic's scientific research platform. His instructions were direct: find all publicly available ultraviolet survey data, standardize it, combine it, and fill in the blanks.

Ultraviolet light cannot reach Earth's surface because the ozone layer blocks it, which means all ultraviolet astronomy depends on space telescopes. Over the past fifty years, multiple satellites have collected fragments of the ultraviolet sky. NASA's Galaxy Evolution Explorer, or GALEX, was the workhorse, capturing roughly 38,000 snapshots between 2003 and 2013 that covered about two-thirds of the sky. But GALEX had blind spots by design—it deliberately avoided bright stars and the dense star fields near the galactic plane to protect its detectors, and its far-ultraviolet channel stopped working after 2009. Other missions from NASA, South Korea, and Europe added pieces to the puzzle, but significant gaps remained. Since GALEX retired, no comparable ultraviolet survey telescope has launched.

Claude Science approached the problem by breaking it into parallel tasks. Multiple sub-agents processed different regions of the sky, standardizing images from different telescopes and different years onto a common coordinate system and brightness scale. This was intricate work—different instruments have different resolutions and sensitivities, and stray light from bright stars had to be separated from the faint ultraviolet background. For the regions where no ultraviolet telescope had ever looked, the team used statistical inference. Claude learned the relationship between ultraviolet brightness and visible light, infrared, and radio wavelengths across the two-thirds of the sky that had been observed, then applied that mathematical relationship to the gaps. The team also layered in estimates of ultraviolet light from more than 100 million individual stars using data from the European Space Agency's Gaia satellite. Crucially, the technical documentation makes clear that this gap-filling used ordinary regression—an ensemble of decision trees—not a language model or image generator. The inferred regions cannot reveal new discoveries and should not be treated as independent observations. Their effective resolution is roughly half a degree to one degree; finer details are not real.

But the AI did not catch everything. When Ménard reviewed the finished map one night, he noticed faint circular traces scattered across the darkest regions—some slightly brighter than their surroundings, some slightly dimmer. These were not celestial objects. Each GALEX exposure covers a circular field about 1.2 degrees across, and the images contained uneven atmospheric ultraviolet airglow from Earth. When this airglow was not fully subtracted during processing, it left circular imprints on the map. Claude had flagged this risk early and assigned agents to conduct two rounds of review, but neither caught it. Ménard described the problem in a single sentence: "I can see individual disks—imprints from single observations. Can you fix it?" Claude traced the residual atmospheric airglow and reprocessed all 38,000 observations. After several hours of computation, the circular artifacts vanished.

The final map went through more than ten iterations. It combines far-ultraviolet and near-ultraviolet data into a single all-sky image, with each pixel marked as either "observed" or "predicted" and accompanied by uncertainty estimates. Ménard emphasizes on the map's webpage that complete does not mean observed. Researchers and students using the inferred regions must preserve the source labels and error ranges.

The ultraviolet sky looks strikingly different from what we see in visible light. The galactic center, brilliant in visible wavelengths, becomes one of the darkest regions in ultraviolet because dust blocks ultraviolet light far more effectively. Star-forming regions like Orion, Scorpius, and the Magellanic Clouds glow with brilliant blue light. Dust clouds around young stars, bubble-like rings blown out by stellar explosions, and delicate filamentary dust structures illuminated by starlight all appear clearly on the map.

Ménard sees the project as a window into a new role for artificial intelligence in science. The AI handled the grinding foundational work—data collection, calibration, computation—while the human researcher made the critical decisions: spotting anomalies, determining how to fix them, and setting the research direction. In his acknowledgments, Ménard wrote that the map resulted from the patient labor of many agents within Claude Science, and that his own contribution was "limited to guiding them along the way." He also noted that nearly every scientific field has a backlog of similar projects—work that could help students and researchers but never gets done because it is too time-consuming. AI is now taking on this foundational engineering that humans lack time for but everyone benefits from. Anthropic has released the map publicly under a Creative Commons license, with the processing code available under an MIT license. It will serve as a reference resource for astronomy until NASA's next-generation ultraviolet survey telescope, UVEX, launches in 2030.