Astronomers converted 116 radio observations spanning 27 years into the highest-definition continuous reconstruction of a black hole jet. Using an AI model named Kine, researchers mapped the blazar 3C 345, revealing bright plasma components appearing to travel at 10 to 13 times the speed of light.
When astronomers look deep into space, they rarely get to watch cosmic events unfold in real time. Instead, they capture snapshots across decades and attempt to piece the history together. A research team tackled this limitation by combining 27 years of archival data with machine learning, yielding a continuous video reconstruction that provides a resolution four times higher than that of any individual image for relativistic jets captured by decades of observations.
Reconstructing Blazar 3C 345 With the Kine Neural Network
The target of the study is a blazar cataloged as 3C 345, situated in the constellation Hercules at a redshift of 0.593. Blazars are powered by supermassive black holes at the center of distant galaxies that shoot out massive, relativistic streams of gas charged with X-rays and gamma rays. Because the jet of 3C 345 points within 3 to 6.8 degrees of our line of sight, its motion appears compressed, creating an optical illusion of superluminal travel.
Between 1995 and 2022, astronomers observed the jet 116 times. These epochs came primarily from the MOJAVE monitoring program operating at 15 gigahertz, utilizing the Very Long Baseline Array (VLBA)—a network of 10 radio antennas spread across the United States. While the VLBA functions together as a virtual telescope thousands of kilometers wide, its limited number of antennas creates incomplete spatial sampling. (VLBA had two programs that caught views of the jet, called BEAM-ME and MOJAVE, which collectively followed hundreds of blazar sources.)
Traditional processing analyzes each observation independently, which can introduce flickering and artifacts between frames. To solve this, the research team deployed an AI neural network called Kine, described in an Aug. 26 study in the journal Nature. Kine processes observations across time while learning spatial and temporal correlations.
Marianna Foschi, a postdoctoral researcher at Caltech, is the study first author.
Resolution and Challenging the Shock Model
Validation tests demonstrated that the simultaneous reconstruction achieved an average effective resolution of approximately 113 microarcseconds, roughly 4.2 times finer than the nominal 475-microarcsecond beam of the array. Furthermore, the total dynamic range reached about 500,000, roughly 140 times the conventional CLEAN result for this particular dataset produced by standard processing.
This increase in detail allowed the team to measure plasma speeds directly across the jet rather than tracking only selected Gaussian components. The results surprised the scientific team. The brightest components in the jet were observed traveling at 10 to 13 times the speed of light, while the surrounding bulk gas zoomed at 9 to 12 times light speed in the same region.
The higher quality of our video reconstruction enabled a detailed measurement of the plasma velocity in the jet,
study first author Marianna Foschi, a postdoctoral researcher at Caltech, told Live Science in an email.
Although these figures describe projected apparent velocities rather than local speeds through space—meaning no physical matter actually outruns light locally—the close speeds between the bright knots and the surrounding fluid challenge long-held assumptions. This is unexpected because the general consensus is that these bright components are shock perturbations moving through the plasma, and as such they should have a higher velocity compared to the surrounding fluid,
Foschi said. Our work does not invalidate the shock model in general, but it puts it into question, at least in the case of this specific source.
Applying Physics-Constrained AI to Future Astronomy
The developers emphasize that Kine is a physics-constrained reconstruction tool rather than a text-to-video generator. The network fits space and time directly to interferometric visibilities, relying on real radio data while modeling intermediate frames that sit up to 5.7 months away from an actual observation.
Several study co-authors had previous experience boosting the resolution of distant black hole images while working at the Event Horizon Telescope Collaboration, the team behind the first image of a black hole. By releasing their code, observations and supporting products to the public, the researchers hope to apply the algorithm to other variable celestial targets.
We believe this method will drastically change the way jet dynamics is studied from observations,
Foschi said, adding that the approach enables a precise measurement of the projected velocity at any point in the jet.