The iconic image of the supermassive black hole at the center of M87, sometimes referred to as the “blurry orange donut”, has been officially modified for the first time using machine learning. The new image also shows a larger, darker central region surrounded by bright accumulating gas in the form of a “thin ring”. The team used data obtained by the Event Horizon Telescope (EHT) in 2017 and achieved full resolution of the series for the first time.

In 2017, the EHT collaboration created an “Earth-sized telescope” using a network of seven pre-existing telescopes around the world to collect data on M87. But because it’s not possible to cover the entire surface of the Earth with telescopes, there are gaps in the data, like missing pieces in a mosaic.

“With our new PRIMO machine learning technique, we were able to achieve the maximum resolution of the current array,” says lead author Lia Medeiros of the Institute for Advanced Studies. “Since we won’t be able to closely study black holes, the details of the image play a crucial role in our ability to understand their behavior. The width of the ring in the image is now about two times smaller, which will be a strong constraint on our theoretical models and gravity tests.”

PRIMO, which stands for Principal Component Interferometric Modeling, was developed by EHT members Leah Medeiros (Institute for Advanced Studies), Dimitrios Psaltis (Georgia Tech University), Tod Lauer (NOIRLab) and Ferial Ozel (Georgia Tech University). Their paper, a PRIMO reconstructed image of the M87 black hole, was published today (April 13). Astrophysical Journal Letters.

“PRIMO is a new approach to the difficult task of generating images from EHT observations,” said Lauer. “This provides a way to make up for the missing information about the observed object needed to create an image that can be seen with a single, Earth-sized giant radio telescope.”

PRIMO is based on dictionary learning, a branch of machine learning that allows computers to create rules based on large sets of educational material. For example, if you send a set of different images of bananas to a computer – with enough training – the computer will be able to determine whether an unknown image is a banana. Beyond this simple case, the versatility of machine learning has been proven in many ways, from creating Renaissance-style artworks to completing an unfinished Beethoven piece. So how can machines help scientists image a black hole? The research group answered exactly that question.

Computers using PRIMO analyzed more than 30,000 high-quality simulated images of gas-accumulating black holes. The simulation ensemble covered a wide range of models of how a black hole accumulates matter, looking for common patterns in the structure of the images. The different structure models were sorted by how often they occurred in the simulation and then blended to provide a highly accurate representation of the EHT observations while providing a highly accurate estimate of the missing structure in the images. The article about the algorithm itself has been published Astrophysical Journal February 3, 2023.

“We use physics to fill in missing data fields with machine learning in a way that has never been done before,” Medeiros added. “This could have important implications for interferometry, which plays an important role in fields from exoplanets to medicine.”