astronomy/ai in a supporting role/The Astrophysical Journal 2026 · v2
Keck spectra pin down distances for six gravitational lenses found by a neural network
Astronomers took near-infrared spectra of strong gravitational lenses and measured the distances of their background galaxies. The systems came from a candidate list assembled earlier by a neural network sifting through survey images.
spectrum · one line per step, placed by what the step does · bright lines used AI
DESI Strong Lens Foundry III: Keck Spectroscopy for Strong Lenses Discovered Using Residual Neural Networks
The Astrophysical Journal, 2026
doi:10.3847/1538-4357/ae4a9a · record aix-00031 v2 · checked 2026-10-07
- AI was for
- Classification
- Model family
- Convolutional neural network
- Checked by
- Experimental8 tested, 6 worked
- Code
- not reported
AI processed or interpreted data, but the main finding does not rest on it.
What this research was about
A massive galaxy bends the light of anything directly behind it, acting like a lens and smearing the background object into arcs or rings. These strong lenses are useful because the shape of the distorted light depends on how much mass the foreground galaxy holds. To use one, though, you need to know two distances: how far away the lens sits and how far away the lensed object sits. Distance is read from a spectrum, by finding known emission lines and measuring how far they have been stretched towards redder wavelengths, a shift that grows with distance. For faint and very distant background galaxies, the useful lines move out of visible light altogether and into the infrared.
The researchers set out to supply those missing distances. They took systems from a catalogue of lens candidates, had them photographed by the Hubble Space Telescope, and then pointed the Keck telescope's near-infrared spectrograph at eight lensed sources whose distances were hard to measure in visible light. Six yielded a distance. Optical spectra from the DESI survey supplied the distances of the foreground lensing galaxies, so that six systems ended up with both numbers.
Where AI came in
The lens candidates were found by a residual neural network, a type of image-recognition model, applied to grz imaging from the DESI Legacy Imaging Surveys in earlier work by the same group. Survey images run to many millions of galaxies, and lenses are rare, so the network's job was to flag the small number of cutouts whose arcs looked like lensing. That stood in for people scanning images by eye, although human graders still scored the flagged candidates and chose which ones to send to Hubble and then to Keck, mainly on whether Keck could see them.
Nothing was learned or trained in this part of the work. The spectra were reduced with the PypeIt software package, with the object traces picked out by hand, and the distances came from fitting Gaussian curves to pairs of emission lines using standard numerical tools. So the measured distances do not rest on the network; it only decided which patches of sky were worth a telescope's time.
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
Six strong lensing systems, drawn from a candidate catalog built by a residual neural network search of the DESI Legacy Imaging Surveys and imaged by HST, were observed spectroscopically. Eight lensed sources were targeted with Keck NIRES and source redshifts were determined for six, between zs=1.675 and 3.332; the two non-detections had 600 s exposures at airmass of about 1.6. DESI optical spectroscopy supplied lens redshifts, the source redshift of one remaining system at 1.5818, and an additional source redshift within the six systems. Combined, the data give complete lens and source redshifts for six systems.
How AI was used
The lensing systems studied here were selected by a residual neural network (ResNet) applied to DESI Legacy Imaging Surveys imaging in earlier work by the same group, which produced a catalog of strong lens candidates; candidates also carried human visual inspection grades, and a subset was chosen for HST Snapshot imaging and then for Keck follow-up on the basis of visibility from Keck, airmass and arc brightness. No learned model is trained or run in this paper: the near-infrared spectra were reduced with the PypeIt package using manual object tracing and no telluric correction, and redshifts were obtained by fitting a two-Gaussian-plus-constant model to selected emission lines with SciPy's curve_fit.
The shape of the work
Structural · the record, drawn
AI
Neural network search for strong lens candidates
Running a trained model over new data to predict, classify or score. The AI stood in for manual curation.
We conducted searches for strong lensing systems using deep residual neural networks (ResNet) on the Dark Energy Spectroscopic Instrument (DESI) Legacy Imaging Surveyswhere the paper describes this · verbatim
no AI
Human grading and target selection
Reducing a candidate set by filtering or ranking, in a single pass.
the selection criteria for the 8 systems in this paper was primarily based on visibility from Keckwhere the paper describes this · verbatim
no AI
HST snapshot imaging confirmation
Physical execution, by hand or by robot.
51 systems were observed by WFC3 in F140Wwhere the paper describes this · verbatim
no AI
DESI optical spectroscopy of lenses
Obtaining raw data, whether by measurement, download or retrieval.
The spectroscopic redshifts for the lenses from the DESI Strong Lens Secondary Target programwhere the paper describes this · verbatim
no AI
Keck NIRES near-infrared spectroscopy
Physical execution, by hand or by robot.
followed up by Keck NIRES over two half-nightswhere the paper describes this · verbatim
no AI
Spectral reduction with PypeIt
Cleaning, filtering, normalising or labelling data already obtained.
We use PypeIt for spectral reduction.where the paper describes this · verbatim
no AI
Emission-line redshift fitting
Extracting understanding from model behaviour.
Redshift fitting was performed by performing a Gaussian-fit to two emission lines in the spectrum.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.
No model is trained or run in this study. A residual neural network search of the DESI Legacy Imaging Surveys, reported in earlier papers of this series, supplied the candidate systems that are followed up spectroscopically here; the measured redshifts themselves do not depend on the network.
discovered in the DESI Legacy Imaging Surveys using residual neural networkswhere the paper describes this · verbatim
we target eight lensed sources at redshifts difficult to measure in the optical range and determine the source redshifts for sixwhere the paper describes this · verbatim
Hubble Space Telescope data for these systems is hosted on MASTwhere the paper describes this · verbatim
What this paper did not report
Technical · absence is published deliberately
- CodeWhether the code is available is not stated.
- Trained model weightsWhether the trained model is available is not stated.
- ComputeThe hardware or time used is not stated.
- Version of ResNet lens-candidate classifier (DESI Legacy Imaging Surveys search)Which version of the model was used is not stated.
About this article
Record aix-00031, version 2, checked by a person on 2026-10-07. The record describes the paper; it does not assess whether the paper's findings are right. The paper is published under CC-BY-4.0; quotations are at most 25 words. How we work · Report an error