In April 2019, the Event Horizon Telescope (EHT) collaboration showed the first image of a black hole in history. The object is located at the center of galaxy M87, about 55 million light-years from Earth. The ring-shaped structure in the image quickly became one of the most recognizable scientific images of our time.
The image shows a dark "shadow" of the black hole, outlined by the light of hot gas swirling around it. This matched the predictions of Einstein's general theory of relativity and reinforced the long-standing idea that supermassive black holes reside at the centers of galaxies.
However, scientists later disagreed over how accurately the details and structure of the ring were conveyed in the first publication. A new study, published in the journal Nature, rechecked the original data sets using new methods. The results not only confirm the earlier work but also offer a fresh look at the structure of the ring.
Obtaining such an image is extremely difficult. According to the authors, it is comparable to photographing a donut on the Moon from Earth. Achieving the required sharpness would require a telescope the size of a planet. The EHT combined signals from radio telescopes around the world, creating a virtual telescope on a planetary scale.
The first publication involved four independent image-processing teams. All of them obtained a ring-shaped structure, but their fine details differed. Such discrepancies can arise from human and algorithmic biases in image reconstruction. In addition, one of the papers questioned the very ring-like nature of the source.
A separate difficulty is measurement errors. Each radio telescope introduces distortions due to its instruments or Earth's atmosphere. Astronomers usually estimate these errors and correct for them during analysis, but residual errors can lead to incorrect conclusions. To cope with this, the researchers developed a data-combining method called "closure invariants." Instead of pairs of measurements, it uses triples - three telescope measurements at once. This makes it possible to eliminate atmospheric and instrumental errors and obtain cleaner information about the structure of the object.
To process the data, the authors created a machine-learning method called GenDIReCT. To reduce bias, the model was trained on terrestrial images - cats, dogs and cars - so it did not know what a black hole looks like. The method underwent numerous checks, including on real data from other galaxies, such as Centaurus A. After that, the training and testing procedure was fixed so that human biases would not affect the result.
On December 4, 2025, the fully validated method for the first time "looked" at the EHT observations for M87. In all data sets, the ring structure proved stable. Its appearance and diameter corresponded to a black hole with a mass of 6.5 billion solar masses, observed almost "face-on" - as if looking down into the eye of a hurricane.
In a direct comparison with the EHT collaboration's images, the new result matched in diameter and overall orientation. There are differences in fine details - the width of the ring and the brightness along its circumference - but the overall structure is the same. To show that the success does not depend only on machine learning, the scientists also developed a second processing method based on traditional statistics. It also confirmed the structure in the new data.
The authors emphasize that the work provides independent confirmation of the structure of the supermassive black hole at the center of M87. In the future, the method may also be useful in other tasks where data from many sensors are combined to map unknown structures - for example, in medical imaging or in the study of earthquakes.