The first image of a black hole, released in 2019 by the international Event Horizon Telescope (EHT) consortium, was named “Breakthrough of the Year” by the journal Science. The orange “donut,” from combined observations from multiple radio telescopes, revealed the face of a multibillion-solar-mass monster lurking at the center of the galaxy Messier 87 (M87), located 53 million light-years away. Earth, in the constellation of Virgo. A new data analysis, backed by artificial intelligence (AI) techniques, provides a refined picture.

The donut of 2019 now looks more like a simple orange ring. It should be noted at the outset that the black hole itself, from which no light can escape, is only made visible by default, a dark silhouette in the center of the accretion disk of superheated gas which emits powerful radiation as it swirls towards him. This transformation of the image is not the result of new observations, but of data processing using machine learning, an AI technique implemented by Lia Medeiros (Institute for Advanced Studies, IAS, Princeton) and fellow EHT members, and described on April 13 in The Astrophysical Letters.

The 2019 image was taken from observations conducted in 2017 by seven radio telescopes spread across the Earth – one of which is located at the South Pole! By synchronizing them, the EHT had obtained the equivalent of an antenna 9,000 kilometers in diameter, under a process called “very long baseline interferometry”. But this technique, whose precision is sometimes compared to the ability to see from Earth the equivalent of a donut on the surface of the Moon, has some shortcomings: the image offered is not comparable to a photograph such as can take telescopes like Hubble, but to a reconstruction from the data collected by each of the radio telescopes. “We deduce the most probable image, which is very complicated, because many parameters come into play”, specifies Alain Riazuelo, researcher at the Institute of Astrophysics in Paris.

New data representation

It is these gaps that Lia Medeiros and her colleagues have attempted to fill. Missing information was simulated by a branch of machine learning called dictionary learning, which allows computers to generate rules from large bodies of data. In this case, the Primo (Principal-Component Interferometric Modeling) system relied on 30,000 simulations of accretion disks, covering a whole spectrum of models of how matter aggregates near black holes. Nourished by this corpus, Primo proposed a new representation of EHT data to better manage observational gaps.

“We are using physics to fill in missing data areas in an unprecedented way using machine learning,” Lia Medeiros said in a statement. This could have important implications for interferometry, which plays a role in fields ranging from exoplanets to medicine. “The new images thus obtained should make it possible to determine with more precision the mass of the black hole of M87 and the physical parameters which determine its current appearance, IAS asserts in the same press release: “The width of the ring in the image is now smaller by a factor of two, which will constitute an important constraint for our theoretical models and our tests on gravity “in the vicinity of black holes, also recognizes Lia Medeiros.

“The new image more clearly delineates the inner edge of the accretion disk”, notes Alain Riazuelo, who considers this result “promising, awaiting confirmation”. The first image of the black hole Sagittarius A*, located at the center of our galaxy, the Milky Way, published in May 2022 by the EHT, could benefit from the same treatment. But the exercise may be more complicated, because “our” black hole, some 2,000 times less massive than that of M87, sees the matter it sucks in gravitate around it at a much faster rate, which implies very rapid changes in physiognomy, difficult to grasp by terrestrial interferometry.

These two black holes, close enough for one and massive for the other, are currently the only ones made “visible” by this technique. This is why astronomers have in their boxes even more ambitious interferometry projects based on a network of space radio telescopes which would make it possible to change scale once again, adds Alain Riazuelo. The challenge, he warns, will no longer be to wait for the hard drives flown in from the South Pole, as with the EHT, but to find a way to recover a huge volume of data transmitted from space.