astronomy/ai produced the result/The Astrophysical Journal 2026 · v2
Neural networks sift 812,118 quasar spectra for hidden gravitational lenses
Researchers searched DESI DR1 quasar spectra for quasars bending the light of galaxies behind them. One neural network flagged 494 candidates; a second estimated the background galaxy's distance. Seven candidates were graded A after human inspection.
spectrum · one line per step, placed by what the step does · bright lines used AI
Quasars Acting as Strong Lenses Found in DESI DR1
The Astrophysical Journal, 2026
doi:10.3847/1538-4357/ae8014 · record aix-00077 v2 · checked 2026-10-08
- AI was for
- Classification, Property prediction
- Model family
- Convolutional neural network
- Checked by
- Held-out575 tested
- Code
- available
The finding the paper is about came from the AI.
What this research was about
A quasar is the blazing centre of a distant galaxy, powered by matter falling into a huge black hole. Because galaxies carry a great deal of mass, they bend the light of anything lying directly behind them, acting as a lens. When the lens is itself a quasar, the quasar's own glare makes the trick hard to spot. One clue survives in a spectrum, the rainbow of light split up by wavelength: faint emission lines belonging to a more distant galaxy appear on top of the quasar's own lines. Picking those extra lines out of hundreds of thousands of noisy spectra by hand is impractical.
The researchers worked through 812,118 quasar spectra from the first data release of DESI, a survey that records spectra of millions of objects. They looked for cases where a background galaxy's emission lines sat at a greater redshift, meaning a greater distance, than the quasar in front. They also wanted an estimate of how far away that background galaxy was, taken from the [OII] doublet, a closely spaced pair of oxygen lines.
Where AI came in
Two convolutional neural networks were trained from scratch. Networks of this kind learn to recognise patterns in data by example rather than by being given rules. Because real quasar lenses are scarce, the training examples were made artificially: the authors added the light of a real, more distant emission-line galaxy, brightened by a magnification factor, on top of a real quasar spectrum. One network then learned to score each spectrum between 0 and 1 for whether it looked lensed. The other learned to read off the background galaxy's redshift. Training ran in two phases that swapped halves of the sample, so each network only judged spectra it had not seen.
At a score threshold of 0.7 the classifier returned 494 candidates from the full set. The redshift network's estimate then seeded a fit to the [OII] doublet, giving a refined distance and a signal-to-noise figure. The networks stood in for the conventional approach of template fitting and human scanning at the shortlisting stage; the existing Redrock pipeline was run as a comparison for recovering source redshifts. After cuts on signal-to-noise and on the gap between the two redshifts, people inspected what remained and graded the survivors, with seven given grade A.
Written by AIxSci from the checked record below, to give context for readers outside the field. It is not part of the record.
The work
Technical · from the record
The authors searched 812,118 DESI DR1 quasar spectra in the range 0.03≤z≤1.8 for quasars acting as gravitational lenses, identified by emission lines from a background galaxy at higher redshift than the quasar. A convolutional neural network classifier was trained on mock lens spectra made by adding real DESI emission-line-galaxy flux to real quasar spectra, and a second convolutional network estimated the background galaxy's redshift, which was then refined by fitting a double Gaussian to the [OII] doublet. At a score threshold of 0.7 the classifier returned 494 candidates; after signal-to-noise and redshift-separation cuts and visual inspection, 7 were graded A, with four of these also showing Hβ, [OIII] 4959 Å and [OIII] 5007 Å emission. One Grade A candidate, DESI J140.0528-02.3751, had been reported as a candidate in earlier work.
How AI was used
Two convolutional neural networks were trained from scratch on mock lens spectra constructed by summing a real DESI quasar spectrum and the noiseless FastSpec spectrum of a higher-redshift emission-line galaxy scaled by a magnification factor drawn from a normal distribution with mean 4 and standard deviation 2. The classifier used six convolutional layers (three with 50 filters, three with 100 filters) followed by fully connected layers of 30 and 25 nodes, a sigmoid output, binary cross-entropy loss, and the Adam optimizer with exponential learning-rate decay, implemented in TensorFlow with scikit-learn used for splitting and metrics. Training used a 10% lens to 90% non-lens ratio and was carried out in two phases that swapped the training and blind halves of the quasar sample, each phase using a 70/30 train/validation split, so that each network was applied only to spectra it had not seen. A third network with the same architecture was trained as a regression model with a mean squared error loss to predict the background galaxy redshift; its prediction seeded a double-Gaussian fit to the [OII] doublet within Δz=0.1, which supplied a refined redshift and an [OII] signal-to-noise ratio. The DESI Redrock template/PCA pipeline, applied to residuals after subtracting the best-fit quasar model, was run as a comparison method for source redshift recovery. Non-learned steps then applied an SNR ≥ 3 cut and a minimum quasar–galaxy redshift separation of Δz=0.1, followed by visual inspection and grading.
The shape of the work
Structural · the record, drawn
no AI
Select quasar and emission-line-galaxy spectra from DESI DR1
Obtaining raw data, whether by measurement, download or retrieval.
We select 812,118 quasars using the HEALPixel-based redshift catalog from the main DESI DR1 surveywhere the paper describes this · verbatim
no AI
Construct mock lensed spectra as positive training examples
Cleaning, filtering, normalising or labelling data already obtained.
the fluxes of the QSO and ELG are summed to obtain a “mock” lenswhere the paper describes this · verbatim
AI
Train CNN lens/non-lens classifier in two phases
Fitting model parameters, including fine-tuning an existing model. The AI stood in for conventional algorithm.
We then use the Phase 1 training sample and the Phase 2 training sample to take advantage of two CNNswhere the paper describes this · verbatim
AI
Train CNN regression network for source redshift
Fitting model parameters, including fine-tuning an existing model. The AI stood in for conventional algorithm.
We train the regression model only on mock lens systems generated following Equation 1, using the 326,689 QSOs from the Phase 1 training samplewhere the paper describes this · verbatim
AI
Score blind quasar samples with the classifier
Running a trained model over new data to predict, classify or score. The AI stood in for conventional algorithm.
Using a threshold score of 0.7, the CNN predicts 494 candidates out of the 812,118 QSOs in both phases.where the paper describes this · verbatim
AI
Estimate and refine background source redshift
Running a trained model over new data to predict, classify or score. The AI stood in for conventional algorithm.
Starting from the redshift predicted by CNN, we fit a double Gaussian model with a fixed separation of 2.7Åwhere the paper describes this · verbatim
no AI
Apply signal-to-noise and redshift-separation cuts
Reducing a candidate set by filtering or ranking, in a single pass.
We keep only candidates with an SNR ≥ 3.where the paper describes this · verbatim
no AI
Visually inspect and grade surviving candidates
Reducing a candidate set by filtering or ranking, in a single pass.
We visually inspect the remaining candidates and grade them either A or B.where the paper describes this · verbatim
What the record says
Technical · every part carries its own basis
+ in the paper~ our reading− not reported
How to read the quotations. A quotation shows where the paper describes something. It does not quote every value beside it: one passage locates a part of the work, and values without their own quotation are our reading of that passage.
All reported lens candidates were selected by the CNN classifier and their source redshifts estimated by the CNN regression network; the result of the paper is that candidate list
we train a convolutional neural network (CNN) on mock lens systems (positive samples) and unlensed QSO spectra (negative samples)where the paper describes this · verbatim
consisting of 575 lensed systems across both phase test sampleswhere the paper describes this · verbatim
The Pipeline and associated code to recreate plots are freely available here.where the paper describes this · verbatim
What this paper did not report
Technical · absence is published deliberately
- Trained model weightsWhether the trained model is available is not stated.
- DataWhether the data are available is not stated.
- ComputeThe hardware or time used is not stated.
- Version of Classification CNN (six convolutional layers, two fully connected layers)Which version of the model was used is not stated.
- Version of Redshift Finding CNN (regression network)Which version of the model was used is not stated.
- Version of RedrockWhich version of the model was used is not stated.
About this article
Record aix-00077, version 2, checked by a person on 2026-10-08. The record describes the paper; it does not assess whether the paper's findings are right. The paper is published under CC-BY; quotations are at most 25 words. How we work · Report an error