AI Model Finds Seven Quasar Lenses in DESI Data

A machine learning model has combed through the spectra of nearly 800,000 quasars collected by the Dark Energy Spectroscopic Instrument (DESI) and singled out seven systems that behave as gravitational lenses. The work appears in The Astrophysical Journal, with Everett McArthur, a graduate student in astronomy at Ohio State University, as lead author. If follow-up observations hold up, the catalogue of known quasar lenses doubles against what earlier surveys had assembled.

His co-authors are Klaus Honscheid and Claire Lamman, both at Ohio State. The work was funded by the U.S. Department of Energy and the European Union’s Horizon 2020 programme.

What is a quasar gravitational lens?

A quasar is the core of a distant galaxy fed by a supermassive black hole. Infalling matter heats up and radiates so fiercely that quasars shine 10 to 100,000 times brighter than the Milky Way, which is why they stay visible across billions of light-years.

Gravitational lensing is a separate effect. When a massive galaxy sits between us and such a source, its gravity bends the light passing by, and a single background object reaches us as several images, an arc or a ring. For a quasar to be lensed, the two systems have to line up almost exactly along the same sightline. That is why these systems are rare, and why catalogues have held so few of them.

McArthur explains why they are worth hunting for.

“Quasars are like the baby pictures of a supermassive black hole. So exploring how we get from quasars to those black holes is really important.”

The payoff is measurement rather than spectacle. Bent light supplies extra geometry, which lets astronomers pin down the properties of objects that are otherwise too bright to pull apart.

How the model cut 800,000 spectra down to seven

The team searched spectra, not images. The model was trained on a small set of mock lenses, artificial examples of quasar lens systems stitched together from real quasar and background galaxy spectra. Its job was to learn the difference between an ordinary spectrum and an odd one.

The search ran in three stages:

• the model worked through spectra of close to 800,000 quasars from DESI
• it narrowed the field to roughly 200 objects that departed from expectation
• the team reviewed those 200 by hand and kept the seven systems that look most convincingly lensed

All seven candidates sit at least 5 to 6 billion light-years from Earth. Confirmation is expected from space-based instruments, the Hubble Space Telescope first among them.

“What this proves is our architecture was able to parse through a diverse array of quasar spectra in a really significant way.”

Software as the first pair of eyes

This is not the first time software spotted something in the sky before an astronomer did. We wrote earlier about a wandering black hole that a program found rather than a person. That case turned on a single anomaly; here the sieve ran across an entire catalogue and handed back a shortlist short enough for humans to inspect one by one.

Counting lenses is not the whole point. McArthur ties the search to the question of how a galaxy and its central black hole grow together.

“By studying the tight correlation between galaxies and black holes, we could understand why our galaxy is the way that it is and perhaps why our own black hole is sometimes dormant.”

What the study does not settle

The seven systems are candidates, not confirmed lenses. A spectrum carrying traces of two objects need not mean lensing, so each candidate has to be checked against a high-resolution image. Until Hubble or a comparable instrument resolves the separate images, seven is an estimate awaiting a verdict.

The method carries its own limits. The model learned on simulated examples, so its reliability against the real sky cannot be scored before the candidates are either confirmed or dropped. Seven is also far too small a sample to say how many quasar lenses exist overall, and in this paper the final call belonged to human inspection rather than the algorithm.

McArthur himself frames the result as a method to extend.

“You can very well expand this type of study to find many rare anomalies in a spectrum. We’re in an era when science has suddenly become more accessible than ever, and applying AI to astronomy and machine learning methods to big data sets is part of that.”

The next step is concrete and countable: observe the seven targets and find out whether the answer is seven or fewer.