🔭 astrophysics

Photometric Redshift Estimation for Rubin Observatory Data Preview 1 with Redshift Assessment Infrastructure Layers (RAIL)

This paper presents the first systematic analysis of photometric redshifts from Rubin Observatory Data Preview 1 using the RAIL framework, demonstrating that machine-learning-based algorithms achieve performance metrics satisfying LSST Year 1 requirements while highlighting the benefits of adding Euclid near-infrared photometry for high-redshift accuracy.

Original authors: T. Zhang, E. Charles, J. F. Crenshaw, S. J. Schmidt, P. Adari, J. Gschwend, S. Mau, B. Andrews, E. Aubourg, Y. Bains, K. Bechtol, A. Boucaud, D. Boutigny, P. Burchat, J. Chevalier, J. Chiang, H. -F. C

Published 2026-07-08

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: T. Zhang, E. Charles, J. F. Crenshaw, S. J. Schmidt, P. Adari, J. Gschwend, S. Mau, B. Andrews, E. Aubourg, Y. Bains, K. Bechtol, A. Boucaud, D. Boutigny, P. Burchat, J. Chevalier, J. Chiang, H. -F. Chiang, D. Clowe, J. Cohen-Tanugi, C. Combet, A. Connolly, S. Dagoret-Campagne, P. N. Daly, F. Daruich, G. Daubard, J. De Vicente, H. Drass, K. Fanning, E. Gawiser, M. Graham, L. P. Guy, Q. Hang, P. Ingraham, O. Ilbert, M. Jarvis, M. J. Jee, T. Jenness, A. Johnson, C. Juramy-Gilles, S. M. Kahn, J. B. Kalmbach, Y. Kang, A. Kannawadi, L. S. Kelvin, S. Liang, O. Lynn, N. B. Lust, M. Lutfi, A. Malz, R. Mandelbaum, S. Marshall, J. Meyers, M. Migliore, M. Moniez, J. Neveu, J. A. Newman, E. Nourbakhsh, D. Oldag, H. Park, S. Pelesky, A. A. Plazas Malagón, B. Quint, M. Rahman, A. Rasmussen, K. Reil, W. Roby, A. Roodman, C. Roucelle, M. Salvato, B. Sánchez, D. Sanmartim, R. H. Schindler, J. Scora, J. Sebag, N. Sedaghat, I. Sevilla-Noarbe, R. Shirley, A. Shugart, R. Solomon, D. Taranu, G. Thayer, L. Toribio San Cipriano, E. Urbach, Y. Utsumi, W. van Reeven, A. von der Linden, C. W. Walter, W. M. Wood-Vasey, J. Zuntz, LSST Dark Energy Science Collaboration

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

Imagine the universe as a giant, three-dimensional city. To understand how this city is built, astronomers need to know two things about every "building" (galaxy): what it looks like and how far away it is.

In astronomy, "how far away" is measured by redshift. Think of redshift like the pitch of a siren on a passing ambulance. As the ambulance moves away, the sound drops to a lower pitch. Similarly, as galaxies move away from us due to the expansion of the universe, their light stretches out, shifting toward the red end of the spectrum. The redder the light, the farther away the galaxy is.

The problem is that for billions of galaxies, we can't get a perfect "siren" measurement (called spectroscopic redshift) because it requires a lot of time and powerful telescopes. Instead, astronomers have to guess the distance based on the galaxy's color. This guess is called a photometric redshift (or "photo-z").

This paper is a "dress rehearsal" for the Vera C. Rubin Observatory, a massive new telescope that will soon take pictures of the entire southern sky. Before the telescope starts its main 10-year mission, the team needed to test their tools on a small batch of real data called Data Preview 1 (DP1).

Here is what they did, explained simply:

1. The Toolkit: RAIL

The team used a software toolkit called RAIL (Redshift Assessment Infrastructure Layers). Think of RAIL as a "Swiss Army Knife" for astronomers. Instead of using just one method to guess distances, RAIL allows them to try eight different algorithms (mathematical recipes) at once to see which one works best.

The "Template" Chefs: Two of the tools (BPZ and LePhare) work like chefs comparing a new dish to a cookbook. They take the galaxy's colors and compare them to a library of known galaxy "recipes" (templates) to see which one fits best.
The "Machine Learning" Students: The other six tools (like FlexZBoost, kNN, and DNF) are like students who learn by example. They are fed a massive list of galaxies where the distance is already known (the training set). They study the patterns between color and distance until they can guess the distance of a new galaxy they've never seen before.

2. The Test Drive

The team took real images from the Rubin Observatory's test camera and applied these eight tools to them.

The Training Class: They used a "gold standard" list of galaxies in a specific patch of sky (ECDFS) where the distances were already known from other surveys. They used this to teach the machine-learning tools.
The Exam: They then tested the tools on a new set of galaxies to see how accurate the guesses were.

3. The Results: How Good Were the Guesses?

The team found that the tools are working very well, but not perfectly.

The "Good" Zone: For galaxies that are relatively bright and not too far away (redshift less than 1.2), the tools are incredibly accurate. The "scatter" (how much the guesses wiggle around the true value) is about 0.03. Imagine trying to guess the distance to a car; if the car is 100 miles away, you'd be off by only about 3 miles. This is good enough to meet the strict requirements for the Rubin Observatory's future science.
The "Hard" Zone: The tools start to struggle when looking at very faint galaxies or those very far away (redshift greater than 1.2). It's like trying to guess the distance of a tiny, dim firefly in the dark; the guess becomes less reliable.
The Outliers: About 10% of the time, the tools make a "catastrophic" error, guessing a galaxy is close when it's actually far away (or vice versa). This is usually because the galaxy's color is confusing (e.g., a very distant galaxy looks like a nearby one because of how its light is stretched).

4. The Secret Weapon: Infrared Light

The team also tested what happens if they add data from the Euclid space telescope, which sees in infrared light (heat radiation).

The Analogy: Imagine trying to identify a fruit by looking at it in the dark. If you only have a flashlight (visible light), it's hard. But if you also have a thermal camera (infrared), you can see the heat signature and identify it much better.
The Result: Adding this infrared data didn't help much for the nearby galaxies (because we could already see them well), but it significantly improved the guesses for the very distant galaxies (redshift > 1.2).

5. The Big Picture

The paper concludes that the "dress rehearsal" was a success. The RAIL toolkit is ready to handle the massive amount of data the Rubin Observatory will soon collect.

They can now produce a "distance map" for billions of galaxies.
They can group galaxies into "bins" based on distance, which is crucial for studying Dark Energy (the mysterious force pushing the universe apart).
While there is still work to be done to fix the errors for the faintest, most distant galaxies, the foundation is solid.

In short: The team built a sophisticated guessing machine, tested it on a small sample of real universe data, and confirmed that it can accurately map the distances of galaxies well enough to help us understand the fate of the universe.

Technical Summary: Photometric Redshift Estimation for Rubin Observatory Data Preview 1 with Redshift Assessment Infrastructure Layers (RAIL)

Problem Statement
The Vera C. Rubin Observatory's Legacy Survey of Space and Time (LSST) aims to probe dark energy through weak gravitational lensing, large-scale structure, and supernovae. These cosmological probes require precise and accurate photometric redshifts ( $z_{phot}$ ) for billions of galaxies to select samples, define tomographic bins, and infer ensemble redshift distributions ( $n(z)$ ). The Rubin Observatory's Data Preview 1 (DP1), utilizing the LSST Commissioning Camera (LSSTComCam), provides the first real observational data to validate processing pipelines. This paper addresses the challenge of estimating $z_{phot}$ from DP1 data, evaluating the performance of various algorithms, and establishing a baseline for future LSST operations.

Methodology
The study employs the Redshift Assessment Infrastructure Layers (RAIL) framework to apply eight distinct $z_{phot}$ algorithms to DP1 photometry. The analysis focuses on two primary fields:

Extended Chandra Deep Field South (ECDFS): Utilizing deep $ugrizy$ (6-band) coverage.
Rubin_SV_38_7: Utilizing $griz$ (4-band) coverage.

Data Preparation:

Input Photometry: The authors use Gaap1p0 (1" Gaussian Aperture) fluxes from the DP1 coadd object catalog.
Selection: Galaxies are selected with $i$ -band PSF flux signal-to-noise (SNR) > 5 and extended source criteria ( $g_{extendedness} > 0.5$ , $r_{extendedness} > 0.5$ ).
Reference Sample: A reference catalog was constructed in the ECDFS field by cross-matching DP1 objects with spectroscopic ($spec$- $z$ ), grism, and high-quality multiband photo- $z$ data from multiple surveys (e.g., 3D-HST, ASTRODEEP, CANDELS, PRIMUS). This reference set was split into 70% for training and 30% for testing, with a confidence threshold > 0.9.
Validation: An independent validation set was created by cross-matching the SV_38_7 field with DESI Data Release 1 (DR1) spectroscopic samples (BGS, LRG, ELG).
Euclid Integration: A subset of the ECDFS data was cross-matched with Euclid Quick Data Release (Q1) to incorporate Near-Infrared (NIR) $YJH$ photometry.

Algorithms Evaluated:
The study compares two categories of methods:

Template-Fitting: BPZ and LePhare. These compare observed photometry to spectral energy distribution (SED) templates, applying priors and $\chi^2$ minimization.
Machine Learning (ML): FlexZBoost, TPZ, kNN, CMNN, GPz, and DNF. These learn empirical mappings from multiband photometry to reference redshifts.

Workflow:
The authors utilized RAIL_projects for workflow management, testing different "flavors" of configurations (6-band, 4-band, and 6-band + Euclid NIR). Hyperparameters were tuned (e.g., increasing tree counts in TPZ, adjusting basis functions in FlexZBoost) to optimize performance.

Key Contributions

First Systematic Analysis of DP1: This work presents the first systematic application of the RAIL framework to real LSSTComCam data, producing per-galaxy probability density functions (PDFs), point estimates, and ensemble $n(z)$ .
Algorithm Benchmarking: A comprehensive comparison of eight algorithms across different data depths (4-band vs. 6-band) and redshift ranges.
Reference Catalog Construction: The assembly of a high-quality, multi-survey reference catalog in ECDFS specifically tailored for training and validating LSST photo- $z$ pipelines.
Euclid NIR Impact: An assessment of the specific benefit of adding Euclid $YJH$ photometry to LSST optical data.

Results

Performance Metrics:
- Across the 6-band data, the machine-learning algorithms achieved a normalized median absolute deviation (scatter) of $\sigma_{NMAD} \sim 0.03$ and outlier fractions ( $\eta_{0.15}$ ) around 10–15%.
- The overall bias for ML-based algorithms was $< 0.005$ , satisfying the LSST Year 1 (Y1) requirement for redshift bias.
- The scatter for all algorithms was $< 0.1$ , satisfying the LSST Y1 requirement.
- Degradation: Performance degrades significantly at faint magnitudes ( $i > 23$ ) and high redshifts ( $z > 1.2$ ), where training data is sparse.
- Algorithm Specifics: FlexZBoost achieved the highest tomographic binning accuracy (78.1%). Template-fitting methods (BPZ, LePhare) showed larger biases and specific failure modes, such as misidentifying Lyman-break galaxies at high $z$ as low- $z$ Balmer-break galaxies.
Ensemble Distributions:
- In the ECDFS test set, most algorithms successfully recovered the true stacked $n(z)$ distribution.
- In the SV_38_7 field (different depth/filter set), some ML methods exhibited artifacts (e.g., high- $z$ tails, spikes at $z \sim 0.7$ ) due to the mismatch between the ECDFS training set and the SV_38_7 target population. DNF provided the least biased $n(z)$ in this field.
Euclid NIR Impact: Including Euclid $YJH$ photometry significantly reduced bias, scatter, and outlier rates for galaxies at $z > 1.2$ . However, for the bulk of the sample ( $z < 1.2$ ), the improvement was minimal due to the concentration of training data at lower redshifts.
Calibration: Probability Integral Transform (PIT-QQ) plots indicated that while most methods are well-calibrated, some (e.g., CMNN) provide overconfident error bars, and template-fitting methods showed PDF skewing toward high redshifts.

Significance and Claims
The paper claims that the RAIL pipeline is ready for producing photometric redshifts for LSST data, demonstrating "promising initial performance" on real commissioning data. The results validate that:

Machine-learning algorithms can meet LSST Y1 requirements for bias and scatter when trained on representative, high-significance samples.
The pipeline successfully generates both per-galaxy PDFs and ensemble $n(z)$ estimates essential for weak lensing and clustering analyses.
The inclusion of NIR data (Euclid) is beneficial for high-redshift populations, though its impact is limited by the current lack of high- $z$ reference training data.

The authors explicitly note limitations: the training set size and depth limit robustness beyond $z \approx 1.2$ ; some algorithms struggle with non-detections; and hyperparameter optimization was limited. They conclude that while the pipeline is functional, future work must address sample variance, non-detection handling, and further optimization to maximize performance on the full LSST object catalog.

1. The Toolkit: RAIL

2. The Test Drive

3. The Results: How Good Were the Guesses?

4. The Secret Weapon: Infrared Light

5. The Big Picture

More like this