16 new quasars at the end of the reionization unveiled by self-supervised learning

By applying self-supervised contrastive learning and spectral energy distribution fitting to DESI Legacy Survey images, researchers overcame the challenge of foreground contamination to identify and spectroscopically confirm 16 new high-redshift quasars ($z=5.94$–6.45), including three missed by traditional methods, thereby demonstrating a scalable framework for uncovering rare early-universe sources.

The Gist

Here is an explanation of the paper, translated into everyday language with some creative analogies.

The Big Picture: Finding Needles in a Cosmic Haystack

Imagine the early Universe (about 1 billion years after the Big Bang) as a vast, dark ocean. Scattered throughout this ocean are Quasars. These are the "lighthouses" of the cosmos—supermassive black holes eating gas so voraciously that they shine brighter than entire galaxies.

Astronomers want to find these lighthouses to understand how black holes grew so big, so fast. But there's a massive problem: The Haystack.

The sky is filled with billions of stars. Among them are a specific type of cold, dim star called an Ultracool Dwarf (UCD). These dwarfs are like "cosmic impostors." They are red, dim, and look almost exactly like the distant, ancient quasars we are trying to find. In fact, for every one real quasar, there are 2,000 to 10,000 of these impostor dwarfs.

Traditionally, astronomers tried to find quasars by using a "color filter." They looked for objects that were very red in specific ways. But this filter was too strict. It was like looking for a needle in a haystack while wearing sunglasses that only let you see perfectly red needles. If a needle was slightly orange or had a weird shape, you missed it.

The New Approach: Teaching a Computer to "See"

In this paper, the team (led by L.N. Martínez-Ramírez) decided to stop using the old "color filter" rules. Instead, they taught a computer using Self-Supervised Learning.

The Analogy: The "Spot the Difference" Game
Imagine you have a giant box of photos. Most are of trees, but a few are of cars.

• The Old Way (Supervised Learning): You show the computer 1,000 photos of cars and say, "This is a car." Then you show it 1,000 photos of trees and say, "This is a tree." The computer learns to spot cars based only on what you told it. If a car looks weird (like a red truck), the computer might miss it because it wasn't in the training photos.
• The New Way (Self-Supervised Learning): You don't tell the computer what a car is. Instead, you show it millions of photos and say, "Here are two photos. Are they the same object or different?" The computer has to figure out the patterns on its own. It learns that "cars" have wheels and "trees" have branches, even if it's never seen a red truck before.

The team used this method on images from the DESI Legacy Survey (a massive map of the sky). They fed the computer millions of images of "dropouts" (objects that disappear in blue light but appear in red light). The computer learned to group similar-looking objects together without being told which ones were quasars.

The Results: 16 New Lighthouses

The computer created a "map" of these objects. On this map, the real quasars naturally clustered together in a specific neighborhood, separate from the impostor stars.

The team then picked the top candidates from this map and pointed giant telescopes at them to take a "spectral fingerprint" (a detailed rainbow of light).

• The Success Rate: They looked at 40 candidates. 16 of them were real quasars. That's a 45% success rate, which is incredibly high for this type of hunt (usually, you might only get 10-20%).
• The Surprise: All 16 quasars were relatively bright and had some weird, unique features:
  • Some had "narrow" light beams (like a laser pointer instead of a floodlight).
  • Some had very strong chemical signatures that usually get filtered out.
  • Crucially: Three of these 16 quasars would have been completely missed by the old color-filter methods. They were the "red trucks" that the old sunglasses would have ignored.

Why This Matters

1. Breaking the Bias: The old methods were biased toward "normal" looking quasars. This new method found the "weird" ones, giving us a more complete picture of the early Universe.
2. Scalability: This method is like a super-efficient search engine. It can handle the massive amounts of data coming from future telescopes (like the Rubin Observatory or the Euclid satellite) much better than human-designed rules.
3. Black Hole Growth: Finding these bright, early quasars helps scientists understand how black holes grew to be billions of times the mass of our Sun in such a short time.

The Takeaway

The astronomers didn't just find 16 new quasars; they found a better way to look. By letting a computer learn the patterns of the Universe on its own, rather than forcing it to follow rigid human rules, they uncovered a hidden population of cosmic lighthouses that were previously invisible. It's a reminder that sometimes, to find the rarest things in the universe, you have to stop looking for what you expect to find, and start looking for what the data actually shows.

Technical Summary

Here is a detailed technical summary of the paper "16 new quasars at the end of the reionization unveiled by self-supervised learning."

1. Problem Statement

The discovery of luminous quasars at redshifts $z > 6$ is critical for understanding the growth of supermassive black holes (SMBHs) and the state of the intergalactic medium (IGM) during cosmic reionization. However, identifying these objects is exceptionally challenging due to:

• Extreme Scarcity: The number density of $z > 6$ quasars drops exponentially, with fewer than one per comoving Gpc$^3$ at $z \sim 6$.
• Overwhelming Contamination: Foreground ultracool dwarfs (UCDs; M, L, and T dwarfs) outnumber high-redshift quasars by 2–4 orders of magnitude. Their red spectra mimic the "dropout" signature (non-detection in blue filters due to Lyman-$\alpha$ absorption) used to select high-$z$ quasars.
• Selection Biases: Traditional selection methods rely on conservative color-color cuts (e.g., $i-z$ and $z-J$) to minimize contamination. This approach systematically excludes rare populations, such as red quasars, obscured AGN, and objects with unusual emission line properties, leading to an incomplete census of the early universe.
• Limitations of Supervised ML: Existing machine learning approaches often rely on supervised training with labeled catalogs, which inherit the biases of the training data and struggle with rare, underrepresented populations.

2. Methodology

The authors developed a novel, self-supervised framework combining Contrastive Learning (CL) with Spectral Energy Distribution (SED) fitting to overcome these limitations.

A. Data Source and Preselection

• Survey: DESI Legacy Survey Data Release 10 (LS DR10), covering $>20,000$ deg$^2$ with deep optical imaging ($g, r, i, z$).
• Preselection: Instead of strict color cuts, they applied loose morphological and photometric constraints to select $i$-dropout candidates ($z > 5.5$).
  • Required detections in $z$-band and non-detections in $g/r$ bands.
  • Required cross-matching with Near-Infrared (NIR) surveys (VHS, UKIDSS, UHS) and WISE W1 to enable SED fitting.
  • Applied compactness criteria to remove extended galaxies and artifacts, resulting in an initial sample of 165,253 candidates.

B. Self-Supervised Contrastive Learning (CL)

• Input: $10'' \times 10''$cutouts from$g, r, i, z$bands (38$\times$ 38 pixels).
• Architecture: A modified simCLR (Simple Framework for Contrastive Learning of Visual Representations) model.
  • Encoder: Four Convolutional Neural Network (CNN) layers reducing spatial dimensions to a 128-length feature vector.
  • Projection Head: A Multi-Layer Perceptron (MLP) mapping vectors to a 10-dimensional latent space.
  • Training: The model learns by contrasting augmented views of the same source (positive pairs) against different sources (negative pairs) without any labels. Augmentations included flips and rotations.
• Latent Space Analysis: The 128-dimensional embeddings were reduced to 2D using UMAP (Uniform Manifold Approximation and Projection). This revealed distinct clusters:
  • Right Island: Dominated by stars and M/L dwarfs.
  • Left Island: Populated by T dwarfs and high-redshift quasars.
  • Key Finding: The CL model naturally separated quasars from contaminants based on subtle morphological and spectral features present in the optical images, even without explicit NIR input.

C. Candidate Prioritization via SED Fitting

To optimize telescope time, the authors prioritized candidates using SED fitting (via eazy-py):

• Templates: Compared candidates against quasar/AGN templates and contaminant templates (UCDs, stars, galaxies).
• Metric: Used the ratio of $\chi^2$ values ($\chi^2_{BD} / \chi^2_{QSO}$) to distinguish quasars from brown dwarfs.
• Selection: Selected candidates with high quasar likelihood ($\chi^2_{BD} / \chi^2_{QSO} > 1$), photometric redshifts $z_{ml} > 5.5$, and narrow redshift distributions.
• Final Sample: Reduced the list to 1,139 high-priority candidates for spectroscopic follow-up.

3. Key Contributions

• Novel Methodology: First application of self-supervised contrastive learning to optical imaging for high-redshift quasar discovery, demonstrating that it can learn representations superior to traditional color-based selection without labeled training data.
• Bias Reduction: The method successfully recovered quasars with properties typically excluded by traditional cuts, such as red NIR slopes ($z-J > 0.4$) and narrow Ly$\alpha$ emission lines.
• Scalability: The framework is designed to be scalable to future massive surveys (Rubin/LSST, Euclid, Roman).

4. Results

• Spectroscopic Confirmation: The team observed 40 candidates and confirmed 16 new quasars at $z = 5.94 - 6.45$.
  • Success Rate: 45% (16 confirmed out of 40 observed).
  • Contaminants: 22 were identified as UCDs or other contaminants.
  • Literature Overlap: 3 of the confirmed objects were previously known but re-identified; 1 was confirmed by DESI DR1; 12 were entirely new discoveries.
• Physical Properties:
  • Brightness: All 16 are relatively bright ($M_{1450} < -25.5$).
  • Emission Lines:
    • Narrow Ly$\alpha$: Four quasars exhibit narrow Ly$\alpha$ lines (FWHM $\lesssim 2600$ km s$^{-1}$), a feature often missed by standard selections.
    • High Equivalent Widths: $\sim 44\%$ of the sample shows enhanced Ly$\alpha$ + N V + Si II emission (EW $> 100$ Å), significantly higher than the log-normal distribution of known quasars.
    • Red Continuum: Several objects show mild red NIR continua ($z-J > 0.4$), suggesting they are dust-obscured or have intrinsically softer spectra.
  • Missed by Traditional Methods: Three of the 16 quasars would have been missed by traditional color-color selections (e.g., DES, PS1, or SDSS criteria) due to their specific color loci or lack of coverage in specific bands.

5. Significance

• Unveiling Hidden Populations: This work proves that self-supervised learning can uncover rare, atypical quasar populations that are systematically missed by conservative, human-designed selection criteria.
• Efficiency: The combination of CL and SED prioritization achieved a high success rate (45%) with a manageable candidate list, offering a robust path for future large-scale surveys.
• Cosmological Implications: By expanding the census of high-$z$ quasars to include those with narrow lines and red continua, this method provides better constraints on SMBH formation scenarios, potentially revealing faster growth rates or different seed mechanisms in the first billion years of the Universe.
• Future Outlook: The approach is directly applicable to upcoming surveys like Rubin/LSST, Euclid, and Roman, promising to push the redshift frontier further and refine our understanding of the Epoch of Reionization.