AI reveals new details about the Milky Way's monster black hole

A new set of images and measurements of the Milky Way’s central black hole – Sagittarius A* (Sgr A*) – a region in space where gravity is so strong that nothing, not even light, can escape. This analysis suggests Sgr A* spins fast and faces us at a slight tilt.

The fresh view of Sgr A* comes from software that learns patterns in telescope signals that used to be tossed out as too messy.

Back in May 2022, astronomers studied our black hole with the Event Horizon Telescope (EHT), a global network of radio observatories that together act as one giant telescope.

That debut raised big questions about how the system behaves over minutes and hours, and how matter swirls and heats near the point of no return.

Using AI to study Sgr A* spin

The study was led by Michael Janssen of Radboud University, who helped design a machine learning pipeline that handles the Event Horizon Telescope’s complex data.

His team trained an approach called a Bayesian neural network, a type of artificial intelligence that combines data patterns with probability theory to estimate both results and uncertainty.

In the new study, the network points to a high spin for Sgr A*, on the order of 0.8 to 0.9 in dimensionless units, and a viewing angle near 20 to 40 degrees.

The analysis also sets a preferred direction on the sky for the spin axis, in a band between 106 and 137 degrees east of north.

The same work looks forward to near-term upgrades to the global radio array.

The team reports that adding the Africa Millimeter Telescope, a new high-altitude observatory planned for Namibia, would cut some error bars by about a factor of three for certain beyond-standard models.

How EHT and AI measure Sgr A* spin

The Event Horizon Telescope is not a single observatory, it is a network of radio dishes that record waves arriving almost in sync and later combine them to mimic a planet-sized mirror.

The technique is called very long baseline interferometry (VLBI), a method that links widely separated telescopes to create an image with extremely high resolution.

This method is powerful, but it is sensitive to tiny timing errors and air conditions over each dish. The collaboration has shown how tropospheric fluctuations, or shifts in the lower atmosphere caused by water vapor and temperature, and site-by-site differences challenge the calibration of these data and why careful processing is essential.

The new project tackles those headaches by modeling the measurement process directly, rather than only working with final images. That choice lets the network learn from parts of the dataset that older pipelines had to discard.

How the neural network learned

To teach the system what to look for, researchers generated millions of simulated black hole snapshots using general relativistic magnetohydrodynamics, a complex physics framework that describes how magnetic fields and charged gas behave under Einstein’s theory of gravity.

These synthetic scenes were passed through a virtual Event Horizon Telescope to create fake observations that look like the real thing.

The neural network architecture, code named ZINGULARITY, was then trained to map those observables to physical properties such as spin, disk orientation, and the electron temperature parameter (ETP). The full framework documents the Bayesian setup and the steps used to avoid overconfidence.

A key benefit of the Bayesian design is that it reports a range of likely answers, not a single hard number. That matters when the sky is variable and the array captures only sparse snapshots during short nightly windows.

What the new image claims

For Sagittarius A*, the inference favors a fast-spinning hole with a prograde accretion disk, meaning the gas and dust orbit in the same direction as the black hole’s spin.

The model prefers a low inclination relative to our line of sight, which means we are looking mostly down onto the flow rather than edge-on.

Those settings help explain why the central ring looks bright on one side and why the polarization pattern, the orientation of light waves caused by magnetic fields, behaves the way it does in other independent datasets.

The team also notes that the jet, if present, should be faint near the hole, which lines up with the lack of a strong jet detection so far in this source.

The results include a sky orientation for the spin axis that overlaps earlier polarization-based hints. That consistency, across very different methods, gives confidence that the network is keying in on real data features rather than artifacts.

Why some scientists are cautious

“It is very difficult to deal with data from the Event Horizon Telescope. A neural network is ideally suited to solve this problem,” said Janssen, underscoring the problem the method is trying to solve. 

Skeptics worry that a network can latch onto subtle biases if the training library leaves out real-world effects or if the telescope misses key baselines, the distances between paired antennas that define the resolution.

Supporters respond that Bayesian estimates, bootstrapped noise models, and cross-checks with traditional pipelines can catch many of those failure modes.

Black hole spin and Sgr A*

A fast spin, if confirmed, sets how energy can be extracted from the hole’s rotation and how particles are accelerated near the inner edge of the flow.

It also narrows theories for how the Milky Way’s center evolved, including possible past galactic mergers, events where two galaxies collide and their central black holes combine.

Knowing the tilt tells us which parts of the supermassive black hole environment are in view and which are hidden.

That detail helps interpret the flickers and flares seen at other wavelengths and links them back to motions in the ring that radio interferometers can resolve.

The outcome also highlights the value of pairing data-driven tools with solid physics. Networks that explain their confidence and are tested on synthetic observations close the gap between raw measurements and parameters that shape our models.

What comes next for the telescope

The Event Horizon Telescope has already been observed with more stations and at multiple frequencies since the 2017 campaign used here.

Re-running the pipeline on those newer datasets should test whether the same spin and orientation drop out again under different observing conditions.

Hardware improvements are moving in parallel. Bringing in new stations on other continents, and lengthening some baselines, should strengthen image fidelity and reduce the chance that atmospheric quirks steer the answers.

As the library of simulations grows, the training set can expand beyond the usual families of disks and magnetic states. That will give the network a wider menu of possibilities and make its conclusions more resilient.

The study is published in Astronomy & Astrophysics.

Image credit: EHT Collaboration/Janssen et al.

—–

Like what you read? Subscribe to our newsletter for engaging articles, exclusive content, and the latest updates. 

Check us out on EarthSnap, a free app brought to you by Eric Ralls and Earth.com.

—–