AI Search of Zwicky Transient Facility Data Finds Black Hole 30,000 Light-Years Off-Center
University of Maryland researchers reported on July 27, 2026, that an AI-assisted search of Zwicky Transient Facility data identified a dormant supermassive black hole roughly 30,000 light-years from the center of its apparent host galaxy. The study published in The Astrophysical Journal Letters describes the first detection of a quiescent black hole at such a large galactic offset. The key operational change was straightforward but consequential: Instead of examining only galaxy centers, the software searched the full survey for the light pattern produced when a black hole disrupts a star.
The event, designated TDE 2025abcr, briefly made an otherwise invisible object observable. A star passing close to the black hole was pulled apart by tidal forces, and its heated debris generated a flare detectable across multiple wavelengths. Researchers estimate that the black hole has a mass of roughly one million Suns, comparable to the same broad mass class as the black hole at the Milky Way’s center.
The discovery began as a data-filtering problem
The Zwicky Transient Facility at Palomar Observatory surveys the northern sky every two days and records hundreds of thousands of changing celestial sources per night. That cadence is useful for catching short-lived events, but it creates a classification bottleneck: Most detections are not evidence of a displaced black hole, and astronomers cannot manually inspect every alert in time to organize follow-up observations.
The Maryland-led team addressed that constraint with a machine-learning classifier trained to recognize the light-curve behavior associated with tidal disruption events. Unlike searches built around the expectation that supermassive black holes sit at galactic nuclei, this program evaluated candidates regardless of their position within or around a galaxy. It began operating in August 2025 and flagged TDE 2025abcr three months later.
This was not a case of software independently proving that a black hole existed. The classifier performed triage: It reduced a large alert stream to a candidate worth examining with other instruments. Follow-up observations found a blue, ultraviolet-bright flare, broad hydrogen and helium emission, and very soft X-ray emission. Those combined signatures supported the tidal-disruption classification, while the apparent host’s spectrum lacked clear indicators of an active galactic nucleus.
That distinction matters because machine learning in survey astronomy is most useful as part of a system. Wide-field telescopes provide cadence and coverage; classification software identifies unusual behavior; and follow-up instruments supply the spectral and multiwavelength measurements needed to test the initial label. The discovery depended on that complete chain, not on an AI model alone.
The offset is the central measurement
TDE 2025abcr appeared 9.3 kiloparsecs, or approximately 30,000 light-years, in projected distance from the center of the massive galaxy associated with it. Researchers also found no observable point source, dwarf galaxy or nuclear star cluster at the flare’s location. That combination a large offset, a supermassive black-hole estimate and no visible local host is why the team considers this the strongest known case for a wandering black hole.
Its origin remains unresolved. One possibility is that the black hole occupies the nearly stripped core of a smaller galaxy undergoing a merger with the larger system. Another is that interactions among three black holes ejected the least massive member from a galactic center. Neither scenario has been established. Deeper late-time imaging, after the flare fades, could reveal a faint surrounding stellar system and help separate a stripped remnant from an ejected black hole.
Rubin will increase the filtering challenge
The result arrives as the U.S.-funded Vera C. Rubin Observatory begins a much larger time-domain survey. Rubin is collecting about 10 terabytes per night and can generate as many as seven million alerts describing changes in the sky. Its 3,200-megapixel camera captures a new image approximately every 40 seconds, with automated alert brokers responsible for filtering, cross-matching and prioritizing transient sources.
Those figures do not guarantee a specific number of additional wandering-black-hole discoveries. They do show why search architecture matters. A survey can only reveal objects represented in its data and admitted by its selection rules. Removing the galaxy-center requirement expanded the Zwicky search space without requiring a purpose-built black-hole telescope, and TDE 2025abcr appeared within three months.
The durable technical lesson is that survey assumptions can function like instrument limits. Zwicky already had the sky coverage and cadence; the machine-learning pipeline changed which signals were allowed through the filter. Rubin will provide a much larger stream, but extracting comparable discoveries will still depend on classifiers, rapid follow-up capacity and careful confirmation. In this case, the decisive advance was not seeing deeper it was looking in places the previous search logic largely excluded.
By David Whitaker — Associate editor for AMI’s aerospace and drone systems desk, translating flight systems, aircraft programs, spaceflight, and UAV developments into accessible technical stories.
