🔭 astrophysics

Quality Assessment of Spectroscopic Data Reduction Pipelines Using Artificial Intelligence: Scrutinizing Data Release 2 from the DESI Survey

This paper introduces an unsupervised machine learning pipeline that successfully identifies a substantial population of anomalous spectra in DESI Data Release 2 missed by standard diagnostics, thereby providing a scalable and reproducible quality-assurance layer for large-scale spectroscopic surveys.

Original authors: V. Torres-Gomez, J. Suarez-Perez, J. E. Forero-Romero, S. Bailey, A. Kremin, B. Dey, R. P. Nathan, S. Panda, J. Aguilar, S. Ahlen, D. Bianchi, D. Brooks, T. Claybaugh, A. de la Macorra, P. Doel, S. Fe

Published 2026-06-23

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: V. Torres-Gomez, J. Suarez-Perez, J. E. Forero-Romero, S. Bailey, A. Kremin, B. Dey, R. P. Nathan, S. Panda, J. Aguilar, S. Ahlen, D. Bianchi, D. Brooks, T. Claybaugh, A. de la Macorra, P. Doel, S. Ferraro, A. Font-Ribera, E. Gaztanaga, S. Gontcho A Gontcho, G. Gutierrez, K. Honscheid, C. Howlett, R. Joyce, S. Juneau, D. Kirkby, O. Lahav, M. Landriau, L. Le Guillou, M. E. Levi, M. Manera, A. Meisner, R. Miquel, J. Moustakas, S. Nadathur, N. Palanque-Delabrouille, W. J. Percival, F. Prada, I. Perez-Rafols, G. Rossi, L. Samushia, E. Sanchez, D. Schlegel, H. Seo, R. Sharples, J. Silber, G. Tarle, 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

Imagine the DESI survey as a massive, high-speed library that doesn't just read books; it takes a "spectral fingerprint" of over 58 million stars, galaxies, and quasars every month. These fingerprints are called spectra.

The problem? The library is growing so fast that hiring a team of humans to read every single fingerprint to check for errors is impossible. It would take forever, and humans get tired.

This paper introduces a new, automated "smart librarian" that uses Artificial Intelligence (AI) to find the messy, broken, or weird fingerprints so human experts only have to look at the ones that actually need attention.

Here is how their system works, broken down into simple steps:

1. The Setup: A Library of 58 Million Fingerprints

The researchers took data from the DESI Data Release 2, which contains about 58.3 million spectra.

The Challenge: In the past, scientists would look at these spectra one by one. Now, there are too many.
The Solution: They built a pipeline that acts like a sorting machine. It doesn't need to be taught what a "bad" spectrum looks like beforehand (no training data needed). Instead, it learns what "normal" looks like by looking at the data itself.

2. The Method: The "Crowd" and the "Outcasts"

The AI uses two main tricks to sort the data:

Trick #1: UMAP (The Map Maker)
Imagine you have a giant room full of people (the spectra). Some look very similar (like a crowd of people wearing blue jeans), while others look very different (someone in a clown suit).
The AI uses a technique called UMAP to shrink this giant room into a small, 2D map. On this map, people who look alike stand close together, forming a dense crowd. People who look different stand far away, isolated in the corners.
- Crucial Point: The AI does this tile by tile. Think of a "tile" as a single night's worth of observations. The AI asks, "Who looks weird in this specific group?" rather than comparing everyone in the whole universe at once. This is important because a "weird" spectrum in one group might be normal in another.
Trick #2: FoF (The Friend-of-Friends)
Once the map is made, the AI uses a rule called Friends-of-Friends. It says: "If you are standing next to someone, you are friends. If you are standing alone in the middle of nowhere, you are an outlier."
It groups the dense crowd together and flags the lonely people (the "outcasts") for human review.

3. The Results: Finding the "Glitches"

The AI processed all 14,199 observation tiles and found about 1.1 million "candidate" spectra that looked weird.

When the human team went to check a sample of these candidates (about 391 of them), they found:

67% were actually broken. They had real problems, like:
- The "Step" Glitch: The brightness of the light suddenly jumped up or down where two different cameras met (like a photo stitched together poorly).
- The "Ghost" Emission: A bright spike in the red part of the spectrum that wasn't really there (a leftover echo from the sky subtraction).
- The "Negative" Blue: The blue part of the spectrum went below zero, which is physically impossible (another sky subtraction error).
Only 4% were caught by the old system. The standard software that usually checks for errors missed almost all of these broken spectra.

The Big Takeaway: The new AI method found a huge population of "sick" spectra that the old tools completely ignored. It acts like a second pair of eyes that catches things the first pair missed.

4. What About the "Weird" Ones That Aren't Broken?

The researchers found that about 33% of the flagged spectra didn't have any obvious technical errors.

These might be genuine cosmic oddities—stars or galaxies that are just naturally strange and rare.
The paper estimates that out of the total 1.1 million flagged items, roughly 218,000 might be these genuine, unique cosmic objects rather than broken data.

5. Why This Matters

Efficiency: Instead of humans looking at 58 million spectra, they only need to look at the ~1 million flagged ones (and even then, they can prioritize the worst ones).
Reliability: It ensures that the data used to study the expansion of the universe isn't tainted by hidden glitches.
Scalability: This method is fast and can be run every night as new data comes in, making it a perfect "quality control monitor" for the future of astronomy.

In a nutshell: The authors built an AI that learns what a "normal" star looks like on any given night. When it sees a star that looks like a "glitchy photo" or a "weird outlier," it flags it. This saves human time and catches errors that the standard software misses, ensuring the DESI survey's map of the universe is as clean and accurate as possible.

Technical Summary: Quality Assessment of DESI DR2 Spectroscopic Data Using Unsupervised AI

Problem Statement
Large-scale spectroscopic surveys like the Dark Energy Spectroscopic Instrument (DESI) now collect data at a scale (millions of spectra per month) that renders traditional visual inspection for quality assurance (QA) impractical. Systematic effects in data reduction can propagate into biases in redshift estimation and cosmological measurements, while spectral anomalies may also represent genuine astrophysical variability or instrumental failures. Existing automated QA methods often rely on supervised learning requiring labeled training data or standard pipeline quality flags (e.g., ZWARN), which may miss subtle reduction artifacts. There is a need for a scalable, unsupervised method to identify problematic spectra without prior labeling, capable of operating across the entire survey footprint.

Methodology
The authors present an unsupervised pipeline for spectroscopic quality assessment applied to 58,291,334 coadded spectra from DESI Data Release 2 (DR2). The method operates independently on each of the 14,199 survey tiles to produce a tile-level outlier catalog. The pipeline consists of two primary stages:

Dimensionality Reduction (UMAP): The pipeline uses Uniform Manifold Approximation and Projection (UMAP) to project high-dimensional flux vectors (concatenated from the Blue, Red, and Near-Infrared arms, totaling ~7,900 wavelength bins) into a 2D embedding. Crucially, the authors apply no denoising, continuum normalization, or band-specific filtering prior to embedding. This deliberate choice ensures that reduction artifacts (e.g., flux discontinuities, calibration residuals) remain part of the signal to be detected. The UMAP hyperparameters (neighbors $N_n=45$ , minimum distance $M_d=1.0$ , cosine metric) are fixed across all tiles to ensure reproducibility and comparability.
Clustering (Friends-of-Friends): The 2D embeddings are processed using Friends-of-Friends (FoF) clustering. This algorithm groups spectra based on Euclidean distance in the embedding space. Spectra forming small, isolated components (size $< N_{min}=5$ ) or singletons are flagged as candidate outliers.

The approach requires no labeled training data; it learns the distribution of "normal" spectra directly from the local tile population. The pipeline preserves standard DESI identifiers (TILEID, PETAL_LOC, FIBER, TARGETID), enabling direct retrieval and audit of flagged spectra.

Key Results

Outlier Detection: The pipeline identified 1,095,816 candidate outliers across the survey. The mean tile-level outlier fraction is approximately 1.96%.
Program Dependence: Outlier rates vary significantly by observing program:
- Dark Program: ~0.76%
- Bright Program: ~2.36%
- Backup Program: ~7.31%
  This suggests observing conditions and target populations are dominant factors in outlier rates.
Spatial Distribution: Analysis of the focal plane reveals that outlier fractions are not uniform. Elevated rates are concentrated near the edges of the 500-fiber spectrograph blocks (petals), indicating edge-dependent instrumental variations (e.g., focus, throughput, or resolution changes) rather than a smooth global gradient.
Visual Inspection Validation: A visual inspection of 391 randomly selected candidates from the Dark and Bright main-survey programs revealed:
- 66.8% of candidates exhibited identifiable spectral anomalies consistent with known reduction effects (e.g., arm-join flux discontinuities, spurious Z-arm emission, negative blue continuum).
- Only 4.1% of these inspected candidates carried a non-zero ZWARN flag from the standard reduction pipeline.
- This large discrepancy demonstrates that the UMAP+FoF method recovers a substantial population of problematic spectra that standard diagnostics miss.
Genuine Anomalies: Extrapolating to the full main-survey catalog, the authors estimate that approximately 218,000 candidates show no identifiable reduction artifacts and may correspond to genuine atypical spectra.

Significance and Claims
The paper claims that this unsupervised framework provides a complementary quality-assessment layer to existing pipeline diagnostics. By isolating a dense core of typical spectra from isolated components, the method concentrates human review efforts on the spectra most likely to be problematic, rather than replacing human judgment.

The authors emphasize that the pipeline is:

Scalable: It processes tiles independently, making it feasible for survey-scale analysis and suitable for near-real-time QA.
Reproducible: Fixed hyperparameters allow for direct comparison of data quality across successive data releases (e.g., tracking the impact of pipeline fixes like the transition from "Kibo" to "Loa").
Practical: It serves as a monitor for reduction performance, identifying subtle instrumental effects (such as those near spectrograph boundaries) that might otherwise go undetected but could impact cosmological analyses.

The work does not claim to define a global taxonomy of all anomalies or to correct spectra automatically. Instead, it positions the tool as a prioritization mechanism for human review, reducing the inspection burden from the full dataset to a manageable subset of high-probability candidates. The authors note that future work will focus on integrating this into daily operations and extending the method to other multi-object spectroscopic surveys.