DESI DR1 Spectral Anomaly Catalog

Anomaly Explorer

A full-DR1-scale autoencoder anomaly-detection pipeline — 27.5M DESI DR1 science-target spectra scored down to a 1,244-object anomaly-score>3 catalogue (data release v2, 2026-09-03). Photo-z from unsupervised latent vectors (σ_NMAD = 0.028). These are exploratory pipeline candidates, not confirmed detections. Browse the top 1,000 by anomaly score in the table.

How It Works Research Status Model on HuggingFace

Top Anomalies

Showing top 1,000 by anomaly score. Click column headers to sort. Each row links to the Legacy Survey image viewer. Full catalog (195,829 objects) available for download.

Min score: Worst band: 1000 shown

#	Score	RA	Dec	Band	rB	rR	rZ	Image	Full

About This Data Release

What Paper 3 actually is — Paper 3 is now disposed as a data release, not a standalone paper: the DESI DR1 Anomaly-Score Candidate Catalogue (v2, science-only, 2026-09-03). It is a machine-flagged catalogue of 1,244 spectroscopically unusual DESI DR1 science-target objects (anomaly_score>3, threshold chosen by the pre-declared elbow rule over 27,547,223 scanned TARGETIDs), released as reproducible data infrastructure for follow-up characterization — not a purity, novelty, or detection claim. A known-object recovery benchmark against VizieR reference classes (BAL quasars, blazars, cataclysmic variables, Lyα emitters, superluminous-supernova hosts) found no confirmed class recovered at meaningful enrichment (1 BAL-quasar match at 4.2×; 0 elsewhere).
This explorer shows that same v2 catalogue: 1,244 science-target anomaly candidates, of which 569 (45.7%) cross-matched an existing SIMBAD or NED entry and 675 (54.3%) did not. The broader multi-survey candidate set (per-survey native retrains across DESI, SDSS, LAMOST, eROSITA, Planck, Gaia, NEOWISE) is published as a dataset at HuggingFace. These are pipeline candidates, not confirmed detections. Click any row to view the Legacy Survey image, read AI analysis notes, and add your own review comments.

Pipeline characteristics (22.5M-spectra autoencoder)
Photo-z from latent vectors: σ_NMAD = 0.028 (R² = 0.79) with zero redshift supervision — the autoencoder spontaneously learned spectral features correlated with redshift. Redshift neuron (lat_067): A single latent dimension that emergently encodes redshift, the strongest individual predictor of spectroscopic z. UMAP clustering: Two distinct populations — 247K B-band noise cluster + 2,575 red-anomaly cluster with unusual spectra. Candidate rate uniform: ~1% across the footprint (Spearman r = 0.03 with depth), consistent with astrophysical candidates rather than depth artifacts. High-z candidates: the gold tier includes candidate z>6 objects flagged for Gunn-Peterson-like troughs — exploratory candidates for follow-up, not confirmed reionization-era QSO discoveries.

27.5M
Science-Target Spectra Scanned

1,244
Anomaly Candidates (score>3)

569

SIMBAD/NED Matched

675

Uncataloged (no SIMBAD/NED)

0.028

σ_NMAD Photo-z

Descriptive Taxonomy Families

Descriptive taxonomy families (25 UMAP/HDBSCAN clusters rolled up into 8 families over the 675 SIMBAD/NED-unmatched objects; sizes sum to 675):

Family	N
0 — low-anomaly-score, main/dark	302
1 — elevated-anomaly-score, main/dark	87
2 — elevated-anomaly-score, sv3/bright	71
3 — extreme-anomaly-score, sv3/bright	61
4 — low-anomaly-score, sv3/bright	44
5 — low-anomaly-score, sv1/other	38
6 — high-anomaly-score, main/dark	36
7 — high-anomaly-score, sv3/bright	36

Scientific Insights

Eight genuinely novel findings from the enhanced 22.5M DESI DR1 catalog — not just "we ran an autoencoder," but concrete scientific contributions including a direct improvement to the flagship f_NL bounce prediction.

The "Redshift Neuron"

Latent dimension 067 has 6× the importance of any other dimension for predicting redshift. The autoencoder spontaneously learned to encode spectral shift without ever seeing redshift labels — emergent representation of a physical property.

Unsupervised Photo-z (σ_NMAD = 0.028)

A simple MLP on the 128-dim latent vectors predicts spectroscopic redshift with R² = 0.79 and 7.7% outlier fraction — competitive with purpose-built photo-z codes using broadband photometry.

"Correctly Classified but Spectrally Anomalous"

2,575 objects where DESI’s pipeline is confident (Δχ² = 963) yet the autoencoder flags unusual features — genuine spectral structure beyond standard templates.

16 NEOWISE IR-Variable Anomalies

16 anomalies are BOTH spectrally anomalous AND infrared-variable (NEOWISE 10yr). A z=5.65 QSO varies by 5.5 magnitudes in W2 — extreme AGN activity in the reionization era. These variable sources are prime candidates for multi-epoch follow-up.

675 Uncataloged Objects

675 of 1,244 v2 anomaly candidates (54.3%) are in NEITHER SIMBAD nor NED. Grouped by PCA→UMAP→HDBSCAN clustering into 25 clusters rolling up to 8 descriptive families (largest: 302 low-anomaly-score, main/dark). A VizieR reference-class recovery benchmark found no confirmed class recovered at meaningful enrichment. Known to DESI, unknown to the astronomical community — concrete targets for follow-up.

Gold Anomalies Cluster in Latent Space

The 83 gold anomalies are 2.2× more clustered than random objects in the 128-dim latent space — confirming a coherent spectral population, not random noise.

Autoencoder as Survey Quality Probe

Anomaly score correlates with SNR (Spearman ρ = −0.89). The autoencoder unintentionally functions as an independent data quality metric — a new tool for spectroscopic survey validation.

+7.93% f_NL Improvement via 5-Tracer Multi-Tracer

5-tracer anomaly-optimized Fisher forecast yields σ(f_NL) = 11.71 vs 12.72 standard multi-tracer (+7.93% improvement). DESI alone contributes 6.1% improvement. Latent-space selection of anomalous objects as high-bias tracers directly strengthens the flagship bounce prediction: spectral anomalies are not just curiosities but observationally useful for testing bounce cosmology via the galaxy bispectrum.

Sky Distribution

All 1,000 top-scored anomalies plotted by RA/Dec. Color indicates anomaly score (yellow = highest, blue = threshold).

How Anomaly Detection Works

What is an anomaly? A spectral autoencoder is a neural network trained to compress and reconstruct normal DESI spectra (stars, galaxies, quasars). When it encounters a spectrum that doesn’t match any learned pattern, the reconstruction is poor — producing a high residual. Objects with total residual (anomaly score) above 5.0 are flagged. These are spectra the model literally “doesn’t know what to do with.”

What the score means: The anomaly score is the sum of reconstruction errors across DESI’s three spectrograph arms. Higher = more unusual. The score tiers are:

5 – 7Mildly unusual — spectrum slightly off from nearest template (151K objects) 7 – 10Moderately anomalous — noticeable spectral features not in training set (34K objects) 10 – 15Highly anomalous — spectrum dramatically different from all known classes (9.8K objects) 15+Extreme outlier — nothing in the training set remotely resembles this spectrum (101 objects)

Column Definitions & Glossary

Table Columns

Score: Total reconstruction error across all three spectrograph arms (B + R + Z). Higher = more anomalous.
RA: Right Ascension (degrees, 0–360). East-west position on the sky in the ICRS coordinate system.
Dec: Declination (degrees, -90 to +90). North-south position on the sky.
Band: Which spectrograph arm has the largest residual: B (blue, 3600–5800Å), R (red, 5760–7620Å), or Z (near-IR, 7520–9824Å).
rB: Reconstruction error in the B (blue) arm. High rB = anomalous blue-end features (e.g. unusual emission lines, UV excess).
rR: Reconstruction error in the R (red) arm. High rR = anomalous mid-optical features (e.g. unusual continuum, absorption).
rZ: Reconstruction error in the Z (near-infrared) arm. High rZ = anomalous near-IR features (e.g. high-redshift emission shifted into IR).
TID: DESI TARGETID — unique identifier for this object in the DESI DR1 catalog.

Astronomy Terms

AGN: Active Galactic Nucleus — a supermassive black hole at a galaxy’s center actively accreting matter, producing bright emission across the spectrum.
QSO: Quasi-Stellar Object (Quasar) — an extremely luminous AGN, often at high redshift (z > 1). Key tracer for large-scale structure measurements.
Near-IR: Near-Infrared — wavelengths just beyond visible red light (~7000–10000Å in the Z-band). High-redshift features shift into this range.
High-z: High redshift — objects at great cosmological distances (z > 1.5), seen as they were billions of years ago.
BAL: Broad Absorption Line — a QSO showing wide absorption troughs from high-velocity outflows. Rare (~10% of QSOs) and often missed by pipelines.
PSF: Point Spread Function — the image of a point source (star or distant QSO). “PSF morphology” means it looks like a point, not an extended galaxy.
REX: Round Exponential — a Legacy Survey morphology classification for a small, round, slightly extended source.
SER: Sérsic profile — a Legacy Survey classification for galaxies fit with a Sérsic surface brightness profile.
SIMBAD: Set of Identifications, Measurements and Bibliography for Astronomical Data — the most comprehensive database of known astronomical objects (CDS, Strasbourg).
NED: NASA/IPAC Extragalactic Database — a database focused on extragalactic objects (galaxies, QSOs, clusters).
f_NL: The amplitude of primordial non-Gaussianity — a key parameter for distinguishing between the Big Bounce and inflation.

Cross-Reference Status

How do we know these are previously unidentified? We cross-match anomaly positions against multiple astronomical databases. An object NOT found in any of these catalogs is a strong candidate for being genuinely new.

Database	What it contains	Objects	Checked?	Matches
SIMBAD	Most comprehensive catalog of identified astronomical objects	~17M	Top 10,000	21/10,000 (0.2%) — 99.8% absent
NED	Extragalactic objects (galaxies, QSOs, clusters)	~400M	Top 10,000	1,270/10,000 (12.7%) — 87.3% absent
Gaia DR3	1.8 billion stars with astrometry & photometry	~1.8B	Top 1,000	6/1,000 (0.6%) — only 1 confirmed Galactic star
SDSS DR18	Sloan Digital Sky Survey — spectra + photometry	~2.3M spectra	77,905 anomalies	Native BigAE rescore complete (3.4% anomaly rate, domain-shift scores)
AllWISE	750M infrared sources — photometric detection catalog	~750M	Top 1,000	15/1,000 (1.5%) — 98.5% have no IR counterpart
Milliquas v8	Comprehensive QSO catalog — all known quasars	~1M	Top 1,000	0/1,000 (0%) — ZERO are known QSOs
Liang+2023 EDR anomalies	Prior DESI EDR autoencoder anomaly catalog	~250K	Pending	Catalog not published as downloadable file
Nicolaou+2026 EDR anomalies	Prior DESI EDR VAE anomaly catalog	~208K	Pending	Catalog not published as downloadable file

Current status: 6 major databases cross-matched, representing over 3 billion cataloged objects. SIMBAD: 0.2% matched. NED: 12.7%. AllWISE: 1.5%. Milliquas: 0%. Gaia: 0.6% (1 star). SDSS DR18: 77,905 anomalies (native BigAE rescore complete). On the v2 science-only catalogue, 675 of 1,244 (54.3%) are in neither SIMBAD nor NED — genuinely uncataloged. Grouped into 8 descriptive taxonomy families (see table above).

Prior Work & Attribution

Prior work: Autoencoder anomaly detection on DESI was pioneered by Liang et al. (2023) on ~250K EDR spectra and Nicolaou et al. (2026, MNRAS, 46 co-authors) on ~208K EDR spectra. This catalog extends their approach by ~90x in scale to the full DR1 release. Both teams must be cited in any publication using this catalog.

Anomaly #1

Legacy Survey