🔭 astrophysics

DESI Strong Lens Foundry III: Keck Spectroscopy for Strong Lenses Discovered Using Residual Neural Networks

This paper presents Keck NIRES and DESI spectroscopic follow-up of strong lenses discovered by Residual Neural Networks in the DESI Legacy Imaging Surveys, successfully determining complete lens and source redshifts for six systems to enable physical parameter extraction and refine future automated lens searches.

Original authors: Shrihan Agarwal, Xiaosheng Huang, William Sheu, Christopher J. Storfer, Marcos Tamargo-Arizmendi, Suchitoto Tabares-Tarquinio, D. J. Schlegel, G. Aldering, A. Bolton, A. Cikota, Arjun Dey, A. Filipp

Published 2026-04-21

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Shrihan Agarwal, Xiaosheng Huang, William Sheu, Christopher J. Storfer, Marcos Tamargo-Arizmendi, Suchitoto Tabares-Tarquinio, D. J. Schlegel, G. Aldering, A. Bolton, A. Cikota, Arjun Dey, A. Filipp, E. Jullo, K. J. Kwon, S. Perlmutter, Y. Shu, E. Sukay, N. Suzuki, J. Aguilar, S. Ahlen, S. BenZvi, D. Brooks, T. Claybaugh, P. Doel, J. E. Forero-Romero, E. Gaztañaga, S. Gontcho A Gontcho, G. Gutierrez, K. Honscheid, M. Ishak, S. Juneau, R. Kehoe, T. Kisner, S. E. Koposov, A. Lambert, M. Landriau, L. Le Guillou, A. de la Macorra, A. Meisner, R. Miquel, J. Moustakas, A. D. Myers, C. Poppett, F. Prada, I. Pérez-Ràfols, G. Rossi, E. Sanchez, M. Schubnell, D. Sprayberry, G. Tarlé, B. A. Weaver, H. Zou

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

The Big Picture: Cosmic Magnifying Glasses

Imagine the universe is a giant, dark room, and we are trying to study tiny, distant fireflies (distant galaxies) that are too faint to see with our naked eyes. Fortunately, nature has placed giant, invisible "magnifying glasses" in front of some of them. These are gravitational lenses.

According to Einstein's theory of gravity, massive objects (like huge clusters of galaxies) warp space itself. When light from a distant background galaxy passes near one of these massive clusters, the light bends. This creates a "funhouse mirror" effect, stretching the background galaxy into arcs, rings, or multiple copies. This not only makes the distant galaxy brighter but also bigger, allowing us to study it in detail.

The Problem: The "Red" Shift

To understand these magnified galaxies, astronomers need to know exactly how far away they are. In astronomy, distance is measured by redshift. As the universe expands, light from distant objects gets stretched, shifting from blue to red. The farther away the object, the redder (and more "infrared") the light becomes.

Here was the problem for the scientists in this paper:

They found about 3,500 potential cosmic magnifying glasses using a super-smart computer program (a Residual Neural Network, or "ResNet") that scanned millions of images from the DESI survey.
They confirmed 51 of these were real using the Hubble Space Telescope.
However, for about 30% of these, the background galaxies were so far away that their light had shifted so far into the infrared spectrum that it was invisible to the optical telescopes (like DESI) that usually do this job. It was like trying to hear a bass drum when you only have a microphone that picks up high-pitched squeaks.

The Solution: The Keck "Night-Vision" Goggles

To solve this, the team used the Keck Telescope in Hawaii, equipped with a special instrument called NIRES (Near-Infrared Echellette Spectrometer).

Think of DESI as a standard pair of glasses that sees well in daylight (visible light). The Keck NIRES instrument is like a pair of high-tech night-vision goggles that can see the "heat" or infrared light that the standard glasses miss.

What They Did (The "Foundry" Process)

The paper is titled "DESI Strong Lens Foundry III." Think of a foundry as a factory where raw materials are melted down and forged into something useful.

The Raw Material: The team started with candidates found by their AI (ResNet) and confirmed by Hubble.
The Forge: They pointed the Keck telescope at 8 of these systems. They didn't just take a picture; they took a spectrum.
- Analogy: If a picture is a photograph of a song, a spectrum is the sheet music. It breaks the light down into its individual notes (colors/wavelengths). By looking at the specific "notes" (emission lines) in the light, they could tell exactly how much the light had been stretched.
The Challenge: Two of the targets were very faint and hard to see because they were observed when the telescope was looking through a lot of atmosphere (high "airmass"), like trying to look through a foggy window. The team had to stop observing them early because the signal was too weak.
The Success: For the other six systems, the "night-vision" goggles worked perfectly. They successfully identified the specific "notes" (like Hydrogen and Oxygen emissions) and calculated the redshift.

The Results: Unlocking the Universe

The team successfully measured the distance (redshift) for six of the most distant background galaxies in their sample.

The Range: These galaxies are incredibly far away, with redshifts between 1.67 and 3.33.
Why it matters:
- Cosmic Map: Knowing the exact distance allows astronomers to build accurate 3D maps of the universe.
- Dark Matter: These lenses act as scales. By knowing the distance to the background galaxy and the shape of the arc, scientists can weigh the invisible Dark Matter in the foreground galaxy cluster.
- Time Travel: Because these galaxies are so far away, we are seeing them as they were billions of years ago. Studying them helps us understand how galaxies formed and evolved in the early universe.

The "Human" Element

The paper also highlights a mix of high-tech and human effort:

The AI: A computer network (ResNet) did the heavy lifting of finding the candidates in the massive data set.
The Humans: Astronomers had to manually check the data, adjust the telescope slits to catch the faint light, and fix software glitches (like a hiccup in the data processing software called "PypeIt") to get the final results.

In a Nutshell

This paper is a success story of teamwork between AI, space telescopes, and ground-based infrared telescopes. The team used a computer to find cosmic magnifying glasses, Hubble to confirm they were real, and the Keck telescope's "night-vision" to measure the distance of the most distant, faint galaxies behind them. They successfully measured the distance to six of these cosmic mirages, providing a crucial key to unlocking the secrets of dark matter and the early universe.

1. Problem Statement

Strong gravitational lensing systems are critical tools for studying dark matter distribution, high-redshift galaxy morphology, and measuring the Hubble constant ( $H_0$ ) via time delays. However, a major bottleneck in utilizing these systems is obtaining accurate spectroscopic redshifts ( $z_s$ ) for the lensed source galaxies.

The Limitation: Many source galaxies in the DESI Legacy Imaging Surveys catalog are at high redshifts ( $z \gtrsim 1.6$ ). At these distances, key optical emission lines (specifically the [O II] $\lambda3727$ doublet) are redshifted beyond the wavelength range of optical spectrographs like the Dark Energy Spectroscopic Instrument (DESI).
The Gap: While DESI successfully measures lens redshifts ( $z_d$ ) for most systems, it fails to secure source redshifts for approximately 30% of the candidates because the necessary spectral features fall into the near-infrared (NIR) regime. Without $z_s$ , physical modeling of the lens mass and cosmological applications are impossible.

2. Methodology

The authors conducted a targeted follow-up campaign using the Keck Near-Infrared Echellette Spectrometer (NIRES) on the 10-meter Keck II telescope to obtain NIR spectra for high-redshift lensed sources.

Target Selection:
- Sources were drawn from a catalog of $\sim3,500$ strong lens candidates discovered using Residual Neural Networks (ResNet) on DESI Legacy Imaging Surveys data.
- The specific 8 systems analyzed in this paper were pre-confirmed as strong lenses with unambiguous arcs by the Hubble Space Telescope (HST) Snapshot program (GO-15867).
- Selection criteria included visibility from Keck, airmass constraints, and the brightness of the lensed arcs.
Observations:
- Observations were conducted over two half-nights (Nov 13, 2022, and Jan 10, 2023).
- Slit Orientation: Slits were aligned along the brightest star-forming "knots" within the lensed arcs to maximize signal-to-noise ratio (SNR).
- Exposure Strategy: An ABBA dither pattern was used to subtract sky background and correlated noise. Most targets received 1200–3600 seconds of integration; two targets received only 600 seconds due to high airmass ( $\approx 1.6$ ) and poor conditions.
Data Reduction:
- Pipeline: The authors used PypeIt, an open-source Python package for spectroscopic reduction.
- Customization: Due to the faintness of the sources and the nature of arc extraction, the authors modified the standard PypeIt workflow:
  - Manual Extraction: Automatic object detection often failed for faint arcs; manual tracing of the slit was employed to extract 1D and 2D spectra.
  - Telluric Correction: Omitted due to low SNR, as it introduced pathological noise behavior without significantly impacting redshift determination based on emission lines.
  - Coaddition: Spectra from multiple dithers and multiple lensed images of the same source were coadded to improve SNR.
Redshift Fitting:
- Redshifts were determined by fitting Gaussian profiles to two distinct emission lines (e.g., H $\alpha$ , [O III], [O II], H $\beta$ ).
- The fitting model included parameters for redshift ( $z$ ), amplitudes, standard deviations, and a continuum normalization constant.

3. Key Contributions

First NIR Spectroscopic Follow-up: This is the first paper in the "DESI Strong Lens Foundry" series to present NIR spectroscopic results for HST-confirmed lenses, specifically targeting the high-redshift gap left by optical surveys.
Pipeline Adaptation: The paper details a robust workflow for extracting spectra from faint, extended lensed arcs using PypeIt, addressing specific challenges like manual trace definition and the suppression of automatic extraction errors.
Validation of ML Candidates: The study validates the efficacy of ResNet-based discovery methods by confirming that candidates with lower human inspection scores (Grade C, score 2.5) still yield high-quality spectroscopic detections.

4. Results

The authors successfully secured source redshifts for 6 out of 8 observed systems. The two failures were attributed to short exposure times (600s) and high airmass.

Redshift Range: The measured source redshifts ( $z_s$ ) span from 1.675 to 3.332.
Specific Findings:
- DESI J094.5639+50.1853: Achieved the highest redshift in the sample, $z_s = 3.33185$ , with clear detections of [O II], [O III], and H $\beta$ .
- DESI J006.3643+10.1853: Identified a complex system with two distinct lensed sources: one at $z_s = 2.39688$ and a second reddish arc at $z_s = 1.3143$ .
- DESI J154.5307-00.1368: Detected a velocity difference of 164.2 km/s between two star-forming knots within the same arc ( $z_{s,1} = 1.73885$ vs. $z_{s,2} = 1.73735$ ), suggesting galactic rotation or distinct knots rather than multiple images of a single knot.
- DESI J023.0157-16.0040: Keck failed to detect a redshift (due to short exposure), but DESI successfully measured $z_s = 1.5818$ in a subsequent observation, demonstrating the complementary nature of optical and NIR surveys at the redshift boundary ( $z \approx 1.6$ ).
Precision: Redshift uncertainties are extremely low, ranging from $O(10^{-4})$ to $O(10^{-5})$ , limited primarily by systematic wavelength calibration errors ( $<0.2$ pixels).
Success Rate: For systems observed with a full ABBA dither pattern (1200s+ exposure), the success rate for obtaining source redshifts was 100%.

5. Significance

Enabling Lens Modeling: The provided redshifts are essential for converting geometric lensing parameters into physical mass profiles, a prerequisite for the subsequent papers in this series (specifically Paper V).
High-Redshift Galaxy Science: The spectra allow for the study of star formation rates (via equivalent widths), metallicity, and galaxy evolution in the $z \sim 1.7 - 3.3$ regime, which is otherwise difficult to probe.
Supernova Searches: The data helps identify ideal systems for lensed supernova searches and estimates supernova rates.
Methodological Benchmark: The paper establishes that NIR spectroscopy (Keck/NIRES) is a highly effective follow-up strategy for DESI-discovered lenses where optical spectroscopy fails, achieving a 100% success rate for sufficiently exposed targets. This validates the ResNet discovery pipeline as a reliable source for high-redshift astrophysical research.