Personal tools
Log in

Skip to content. | Skip to navigation


Istituto italiano di astrofisica - national institute for astrophisics



An international group of scientists, including researchers from the Italian National Institute for Astrophysics, has used artificial intelligence (AI) to measure the size of galaxies as far away as about seven billion light-years from Earth

An artistic view of a convolutional neural network. Credits: Li/Napolitano/Tortora/KiDS-collaboration/ESO

Rome, 25 April 2022 - An international group of scientists, including researchers from the Italian National Institute for Astrophysics, has used artificial intelligence (AI) to measure the size of galaxies as far away as about seven billion light-years from Earth. The researchers - coordinated by Nicola R. Napolitano (full professor at the Sun Yat-sen University in China) - developed a convolutional neural network to determine the structural parameters of galaxies, particularly their size. It is the first time this technique has been applied to ground-based data, demonstrating that AI represents a much faster and more accurate technique. Furthermore, it is the future of large-field surveys from Earth (Rubin/LSST) and space (Euclid), where astronomers will study billions of galaxies. These massive databases, to the construction of which Italian researchers will provide a fundamental contribution thanks to AI, will allow us to explore the physical processes that have guided the evolution of galaxies from the primordial ages of the Universe to today. This study is reported in an Astrophysical Journal article.

"Just as a tailor who has the eye and the experience to determine a person's size to sew him the perfect suit, the astronomer needs to know the shape and size of galaxies. These are crucial information to understand their structure and reconstruct the models to explain their evolutionary history”, says Crescenzo Tortora, a researcher at INAF and author of the article. Tortora has contributed in recent years to the development of these techniques to search for gravitational lenses.

While it may be reasonable to think that a higher mass corresponds to larger sizes, the physical processes regulating the scaling relation between size and mass can vary. For example, astronomers find that the size of more massive galaxies varies significantly with mass. Elliptical galaxies have stellar masses of more than 10 billion solar masses. Moreover, they have changed up to 4 times in the last 10 billion years.

This result has a huge impact on our understanding of galaxy evolution. This variation in size can be explained by galaxy merging, i.e., merging with galaxies with the same mass and size (major merger) or smaller and less massive galaxies (minor merger). This last type of merger seems to be the most accredited process to explain this substantial evolution of the galactic sizes. Therefore, studying galaxies of different masses located in different galactic environments (and therefore more or less close to other galaxies) and at different cosmic epochs is crucial to understanding how these processes occur.

Ongoing observational surveys, such as KiDS, HSC, DES and future ones carried out with the Euclid space telescope and the 8-meter Rubin telescope, will allow us to observe billions of galaxies using fast and effective methods of analysis. The researchers have relied on the help of the eye of an artificial tailor, an artificial neural network, which can find the model that best fits each galaxy.

The first author of the work, the post-doc Rui Li, of the University of the Chinese Academy of Sciences in Beijing, explains: "In the last years, we have developed convolutional artificial neural networks, i.e. software that reproduces the neural connections inside the animal visual cortex, to find gravitational lenses, rare events that occur when light from a distant background galaxy is deflected by the gravitational field of a foreground (lens) galaxy. We identified these galaxies by recognising particular shapes, such as multiple images, arcs and rings, and deformed images of the distant background source. We have extended this type of analysis, developing a neural network capable of reproducing the characteristic parameters of galaxies. This machine learning technique is based on the network's training obtained by providing examples of what it will have to determine, in this case, simulated images of galaxies with known values ​​of some characteristic parameters. And then, we simulated galaxies with analytical models called Sérsic light profiles. Next, we trained the network to obtain some parameters, including the radius that contains half of the total light (which is a proxy of the size of the galaxies), the slope of the light profile, the ellipticity and the full light of the galaxy”.

"We have achieved exceptional performances that only artificial intelligence can allow us to achieve", says Napolitano, project coordinator. “With this work, we have leapt forward, using convolutional neural networks to do something that has not been done before. We took into account that atmospheric seeing, i.e. the blur effect due to atmospheric turbulence that deflects the path of light rays coming from space, systematically varies from galaxy to galaxy in the images collected by our telescopes. Therefore, we trained our neural network to learn the properties of the galaxy under analysis and the most appropriate seeing in the region of the sky in which it is located. In this way, we were able to obtain structural parameters of the galaxies with an accuracy almost comparable to that which would be obtained by taking the same telescope out of the atmosphere. The application to the galaxies observed by KiDS, Kilo Degree Survey, a public survey of the VST telescope, conceived in Naples and built by INAF, and the comparison with the same parameters obtained with a standard technique, further confirmed the goodness of our analysis. This has given us further awareness of our work's use for future surveys".

Tortora adds: "We live in an era that has seen the number of galaxies that telescopes are observing enormous increase, and shortly Euclid and Rubin will provide us with data in the optical and infrared frequencies for millions, if not billions, of galaxies. Our work shows that once the network is trained, we can determine galaxy sizes quickly and accurately. These analyses will provide new information on the galaxies’ structural properties, in clusters, groups or isolated in the Universe, and at different cosmic epochs, to trace the evolutionary processes that led them to be the galaxies they are, from their first cry, until they get old and tired. We are looking forward to our neural network being applied to those data when the Rubin telescope will start observing".

The authors of the work are keen to finish by telling us that their work does not stop there. There will still be a lot to do. Billions of galaxies are waiting for them, and they want to measure their sizes to make the right dress for them. To do so, they will have to measure its size soon. But they did not stop there because, using this automatic technique, they are working to determine galaxy distances, stellar masses and other stellar parameters, and ultimately their dark matter content. So we await these new results with trepidation.

Related journal article:

GAlaxy Light profile convolutional neural NETworks (GaLNets). I. fast and accurate structural parameters for billion galaxy samples”, R. Li, N. R. Napolitano, N. Roy, C. Tortora, F. La Barbera, A. Sonnenfeld, C. Qiu, S. Liu, The Astrophysical Journal.



INAF Press Office - Marco Galliani: +39 335 1778428,

Filed under: ,


Jan 16, 2023

FIRST IMAGE OF A REGION OF THE MILKY WAY FROM THE PEGASUS SURVEY Led by INAF and Macquarie University, a portion of our Galaxy has been imaged in great detail as part of the PEGASUS survey - a radio astronomy project designed to discover more about the Milky Way

Studying the birth of exoplanets with chemistry

Sep 23, 2022

Studying the birth of exoplanets with chemistry A new study led by Elenia Pacetti, PhD student at La Sapienza University and INAF, jointly uses ultra-volatile, volatile, and refractory elements in the atmospheres of giant planets to develop a unified method to shed light on how and where giant planets form. The new work, published in The Astrophysical Journal, paves the road to the exoplanetary studies of the ESA mission Ariel