Astronomers apply machine learning techniques to find quasars from the early universe through vast amounts of data

Machine Learning

AI helps us find cosmic gems hidden in the data ocean

A depth image from the Dark Energy Survey showing the area covered by one of the Dark Energy Camera's individual detectors. Credit: DES Collaboration/NOIRLab/NSF/AURA/M. Zamani

Quasars are extremely luminous galactic nuclei where a great deal of light comes from gas and dust falling into a central supermassive black hole. Due to their extraordinary brightness, these objects can be seen at high redshifts, i.e. at great distances.

A large redshift indicates that quasars are not only far away, but also old in time, and astronomers are interested in studying these ancient objects because they hold clues about the early development of the universe.

High-redshift quasar candidates are first identified by their color (very red), and then we have to look at individual observations of their spectra to confirm that they are high-redshift quasar candidates. However, some high-redshift candidates are sometimes erroneously excluded from further searches because their appearance is distorted by gravitational lensing.

This happens when a massive object, such as a galaxy, sits between us and a distant object: the mass of the galaxy bends space and acts like a magnifying glass, bending the path of light from the distant object and distorting the image of the object.

Although this arrangement has some benefits – gravitational lensing can magnify images of the quasar, making it brighter and easier to detect – it can also deceptively change the quasar's appearance.

Interfering light from stars in intervening lensing galaxies makes quasars appear bluer, while distortions in space-time can make quasars appear fuzzy or multiplied. Both of these effects likely rule them out as quasar candidates.

So a team of astronomers led by Xander Byrne, an astronomer at the University of Cambridge and lead author of the paper announcing the results of this study, Monthly Bulletin of the Royal Astronomical Societyaimed to recover lensed quasars that had been overlooked in previous surveys.

Byrne searched for these lost treasures in the vast data archives of the Dark Energy Survey (DES), conducted using the U.S. Department of Energy-built Dark Energy Camera mounted on the Victor M. Blanco 4-meter telescope at the National Science Foundation's Cerro Tololo Inter-American Observatory, a program of the NSF NOIRLab.

The challenge then was to devise ways to discover these cosmic gems in the vast ocean of data.

The complete DES dataset contains more than 700 million objects. Byrne pared this archive down by comparing the data with images from other surveys to filter out unlikely candidates, such as objects that are likely to be brown dwarfs, which are completely different from quasars in almost every way, yet can look strikingly similar to quasars in images. This process resulted in a much more manageable dataset of 7,438 objects.

Byrne knew he needed to maximize efficiency when searching through the 7,438 objects, but that traditional methods would likely miss the high-redshift lensed quasars he was looking for. “To avoid prematurely ruling out lensed quasars, we applied a contrastive learning algorithm, and it worked beautifully.”

Contrastive learning is a type of artificial intelligence (AI) algorithm that uses a series of decisions to classify each data point into a group according to what it is or isn't. “It may seem like magic, but the algorithm doesn't use any more information than is already in the data,” says Byrne. Machine learning is all about finding what data is useful.

Byrne's decision not to rely on human visual interpretation led him to consider an unsupervised AI process, meaning the algorithm itself, rather than a human, drives the learning process.

Supervised machine learning algorithms are based on so-called ground-level truth defined by a human programmer: for example, the process starts with a description of a cat and proceeds to determine “this is/is not a cat, this is/is not a black cat” etc.

In contrast, unsupervised algorithms do not rely on a human-specified definition as the basis for their decisions. Instead, the algorithm sorts each data point according to its similarity to other data points in the set. In this case, the algorithm might find similarities between images of multiple animals and group them into cats, dogs, giraffes, penguins, etc.

Starting with Byrne's 7,438 objects, an unsupervised algorithm classified the objects into groups. Incorporating geographic similarities, the team called the data groupings “archipelagos.” (The term does not imply spatial proximity between objects; it is their properties that group objects “nearby,” not their location in the sky.)

Within this archipelago, a subset of small “island”-like objects were grouped together as possible quasars. Among those candidates, four stood out like gems in a pile of pebbles.

Using archived data from the Gemini South Telescope, one half of the International Gemini Observatory operated by NSF NOIRLab, Byrne confirmed that three of the four “Quasar Island” candidates are indeed high-redshift quasars. And one of them is very likely the cosmic treasure Byrne hoped to find: a gravitationally lensed high-redshift quasar. The team is now planning additional imaging to confirm the quasar's lensing effect.

“If confirmed, finding one lensed quasar in a sample of four targets would be an astonishingly high success rate! And had this search been conducted using standard search methods, it is highly likely that this gem would have remained hidden.”

Byrne's work is a clever example of how AI can help astronomers mine an increasingly vast treasure trove of data. A massive influx of astronomy data is expected in the coming years, thanks to the five-year Dark Energy Spectrometer survey and the Legacy Surveys and Space and Time, both of which will be run by the Bella C. Rubin Observatory from 2025.

For more information:
Xander Byrne et al., “Quasar Island – Three New z ∼ 6 Quasars, Including Lensed Candidates, Identified by Contrastive Learning.” Monthly Bulletin of the Royal Astronomical Society (2024). DOI: 10.1093/mnras/stae902

Courtesy of NSF's NOIRLab

Quote: Astronomers apply machine learning techniques to find quasars in the early universe amongst vast amounts of data (July 11, 2024) Retrieved July 11, 2024 from

This document is subject to copyright. It may not be reproduced without written permission, except for fair dealing for the purposes of personal study or research. The content is provided for informational purposes only.

Source link

Leave a Reply

Your email address will not be published. Required fields are marked *