Astronomers had only 116 observations of the jet erupting from blazar 3C 345 across 27 years. An AI system turned them into the highest-resolution continuous reconstruction yet—and revealed features appearing to travel 10–13 times faster than light.

Astronomers did not film the jet of 3C 345 continuously for 27 years. They observed it 116 times between 1995 and 2022, then used a neural representation to infer one changing radio image consistent with those irregular measurements.

The work appears in the peer-reviewed Nature paper Video reconstruction of variable VLBI observations with neural fields, led by Marianna Foschi. Its kine algorithm produced a time-continuous, polarimetric reconstruction and a two-dimensional map of apparent motion across the jet.

This is one paper applied to one exceptionally well-monitored blazar. The reconstruction is constrained by real radio interferometry, but frames between observing dates are modelled estimates, not missing photographs recovered from an archive.

The source is aimed almost at us

3C 345 is a blazar at redshift 0.593. Its central supermassive black hole drives a relativistic jet inclined only about 3 to 6.8 degrees from our line of sight. That near alignment makes the jet bright and creates the geometry behind its apparent faster-than-light motion.

The 116 epochs came from the MOJAVE monitoring programme at 15 gigahertz. They were collected with the Very Long Baseline Array, ten radio antennas spread across the United States and operated together as a virtual telescope thousands of kilometres wide.

That method is related to the long-baseline technique behind the Event Horizon Telescope. Space Daily previously described how widely separated observatories can act like one Earth-sized instrument. The gaps between antennas still leave incomplete spatial sampling, so making an image requires careful reconstruction.

Kine fits space and time together

Traditional processing usually reconstructs each observation independently. Small differences in telescope coverage and noise can make the resulting sequence flicker, complicating attempts to distinguish physical motion from image-to-image artefacts. Tracking broad Gaussian components also reduces a changing flow to a few selected features.

Kine instead represents brightness and polarization as a function of right ascension, declination and time. A multilayer neural network predicts those quantities, a forward model converts them into the radio measurements the array should have recorded, and optimization reduces the difference from the actual data.

The team has released code, observations and supporting products. The method is a physics-constrained reconstruction. It is not a text-to-video generator. The network is fitted directly to interferometric visibilities and related observables, while learning correlations across space and time.

The sharpness comes with interpolation

Validation tests gave the simultaneous reconstruction an average effective resolution of about 113 microarcseconds, roughly 4.2 times finer than the nominal 475-microarcsecond beam. Its total dynamic range was about 500,000, around 140 times the conventional CLEAN result for this particular dataset.

Those factors are not universal promises for every radio observation. The authors state that the gains depend on the amount, quality and cadence of the data. Much of the resolution improvement already appeared when kine processed individual epochs, while combining all 116 epochs contributed strongly to the dynamic range.

For optical-flow analysis, the continuous model was sampled at regular times. The most distant interpolated frame lay 5.7 months from an actual observation. Synthetic tests designed to resemble 3C 345 indicated motion-preserving interpolation remained reliable to about six months. Even so, measured epochs and inferred frames are not the same thing.

Nothing travelled faster than light

The bright components appeared to move at 10 to 13 times light speed, while the surrounding bulk flow averaged 9 to 12 times light speed in the same region. These are projected apparent velocities, not local speeds through space, and they do not violate relativity.

A jet travelling close to light speed and pointed almost toward Earth nearly catches up with its own earlier light. The material crosses a real distance between two emissions, but the second signal has a shorter journey to us. The arrival-time interval becomes compressed, making motion across the plane of the sky appear superluminal.

No plasma or information outruns light locally. The paper’s synthetic tests found that kine combined with optical flow could recover apparent speeds up to at least 23 times light speed in data of this quality, above the fastest motion measured in 3C 345. That test addresses whether the algorithm itself would suppress rapid apparent motion.

The bright knots may not be strong shocks

A longstanding interpretation treats compact bright components in relativistic jets as travelling shocks that compress magnetized plasma. Kine let the team compare those pattern speeds with the local flow around them, rather than tracking only the motion of selected knots.

The knots and surrounding plasma had comparable apparent speeds. The reconstructed polarization also lacked the localized increase the authors expected from strong shocks. Together, those findings led them to favour overdense or unusually emissive regions instead of strongly shocked plasma.

That interpretation remains conditional. Some shock conditions can produce similar pattern and flow speeds, and the velocity map assumes that changes in radio emission trace physical plasma motion. The result narrows one explanation rather than eliminating shocks from relativistic jets generally.

The work complements earlier high-resolution jet studies, including Space Daily’s coverage of Centaurus A, but adds a continuous time dimension that a collection of static images cannot supply.

A method as important as the particular jet

The immediate result is a more detailed account of how 3C 345 bends, expels bright structures and transports plasma. The broader proposal is that the same approach can process other long-running VLBI monitoring programmes and short, rapidly changing Event Horizon Telescope observations.

The algorithm does not create extra observations. It uses patterns shared across nearby epochs to estimate a continuous source compatible with the measurements, then validates that process on synthetic data whose true evolution is known.

That distinction is central to reading the result. Kine has turned 116 sparse radio datasets into the highest-resolution continuous reconstruction yet of this jet, not into direct footage of every day across 27 years. With that boundary visible, the reconstruction becomes a new way to measure motion that isolated snapshots could not reveal.