Discovering Hidden Galaxy Mergers: Finding Mergers Missed by Detection Methods

The universe is built by collisions. Every galaxy, including our own Milky Way, grew by merging with others over billions of years. These events trigger star formation, feed supermassive black holes, and drive up to 71% of how galaxies grow in the early universe. When our tools for detecting them have a blind spot, we are getting the history of the universe wrong.

Right now, the James Webb Space Telescope is finding galaxies that are far more massive than our models say should exist. One leading explanation is that we have been systematically undercounting the mergers that built them.

I discovered that a whole class of galaxy collisions erases its own evidence, hiding in telescope data for over 20 years. I identified why they hide, built a machine learning algorithm to find them, and discovered hundreds of real galaxies that every existing tool had missed. Nobody was looking in the right place."

Hi, I am Jared Qin, a Grade 10 student from Victoria, BC.

I compete in astronomy olympiads and have represented Canada internationally, which exposed me to current research and researchers thinking about open problems in the field. One question kept coming up: why do our tools for counting galaxy collisions disagree so much with each other?

Galaxy collisions are the primary mechanism by which the universe builds itself. Every galaxy, including our own Milky Way, grew by merging with others over billions of years. These events trigger star formation, feed supermassive black holes, and shape every structure we can observe. A 2025 study found that mergers drive up to 71% of how galaxies grow in the early universe. So when our detection tools miss an entire class of these events, we are not just miscounting statistics. We are getting the history of the universe wrong.

The specific inspiration came from reading a research paper that mentioned, almost as a footnote, that a population of mergers was being systematically missed by every standard detection method. That line bothered me. Around the same time, news about the James Webb Space Telescope finding galaxies far more massive than our models predicted was everywhere in the astronomy community. The two things connected. Either our physics is wrong, or we are undercounting the mergers that built those galaxies. Nobody had specifically gone after the missing population and asked why it was invisible. That became the question.

Anyone building models of galaxy evolution or trying to explain what JWST is seeing benefits from a more complete merger catalog.

The project has three phases.

I started with the IllustrisTNG cosmological simulation, a detailed computer model of the universe where the complete merger history of every galaxy is known with certainty. This is critical because in real telescope data, you can never know for sure whether a galaxy has merged. The simulation gives you the answer key.

I ran a morphology measurement tool called statmorph on thousands of synthetic galaxy images generated from the simulation. This tool measures how distorted each galaxy looks using statistics like asymmetry, which quantifies how lopsided a galaxy appears, and dozens more. By comparing galaxies that had merged to those that had not, I identified which mergers were completely invisible to every standard detection threshold and studied what physically distinguished them.

This analysis led to the key insight of the project. In a fading merger, the galaxy center settles back to normal first because gravity is strongest there. But faint disturbances persist in the outer regions much longer. I engineered a new feature capturing this: the ratio of outer asymmetry to central asymmetry. When the outside of a galaxy is more disturbed than the inside, a fading merger may be hiding there. Standard tools never compute this ratio.

In Phase 2, I used this insight to train a machine learning classifier on thousands of simulated galaxy images, testing multiple classifier types across many feature combinations with careful controls to prevent data leakage.

In Phase 3, I deployed the trained classifier on a large sample of real galaxies from the Sloan Digital Sky Survey, over 20 years of real telescope data. I then cross-matched the results against Galaxy Zoo, a citizen science project with over 125 million independent human classifications of the same galaxies, to validate the discoveries through a completely separate method.

The central finding is that a systematic blind spot exists in galaxy merger detection. A measurable fraction of all mergers is completely invisible to every standard detection threshold simultaneously. These are not random galaxies. They are preferentially the most massive, most evolved, gas-poor elliptical galaxies in the survey, precisely the population where merger disturbances relax fastest, and detection is hardest.

The key diagnostic is the outer-to-central asymmetry ratio. Hidden merger candidates show dramatically elevated values compared to normal galaxies. The separation between the two populations is statistically large and highly significant, with an effect size of r = 0.526 and p less than 0.001 on the Mann-Whitney U test. This means that in the overwhelming majority of cases, a randomly selected hidden merger candidate has a higher outer-to-central asymmetry ratio than a randomly selected normal galaxy. This is the physical signature of a fading merger remnant where the centre has settled, but the outskirts retain tidal disturbances.

The machine learning classifier achieves strong performance on a held-out test set of galaxies it never saw during training. The area under the ROC curve, a measure of how reliably the classifier separates mergers from non-mergers across all possible thresholds, is 0.898. This means the classifier correctly identifies which of two galaxies is a merger approximately 90% of the time. The detection rate at the operating threshold is 81.3%, with a false positive rate of 16.3%. This represents the highest detection rate achieved on real survey data without requiring spectroscopic observations, surpassing the published benchmark by over 8 percentage points.

Applied to a large sample of real Sloan Digital Sky Survey galaxies, the classifier identified a catalogue of hidden merger candidates that failed every standard detection threshold but scored above the merger probability threshold. These are real galaxies that have been in telescope data for over 20 years without being recognized as mergers.

The Galaxy Zoo validation is the most compelling result. Galaxy Zoo is a citizen science project where human volunteers independently classified the same galaxies with no knowledge of this classifier, no access to the simulation training data, and no connection to the methodology. Cross-matching the hidden candidates against Galaxy Zoo showed that the majority of matched candidates were independently classified as smooth or normal by human volunteers. Two completely different methods, one algorithmic and trained on simulations and one human and based on visual inspection, reached the same conclusion about these galaxies. That independent agreement provides strong evidence that the hidden merger signature is real and consistent across detection approaches.

Physical characterization of the candidate population confirms the theoretical prediction. Over 90% are early-type elliptical or lenticular galaxies. The majority fall on the red sequence, indicating old, quiescent stellar populations. Their outer regions are systematically more disturbed than their centres, exactly the pattern the AO/A ratio was designed to detect.

Every published study that has counted galaxy mergers using standard morphological tools has been working with an incomplete catalog. The missing population is not random. It is concentrated in the most massive, most evolved galaxies, the ones that dominate the high-mass end of the galaxy population and matter most for understanding how the largest structures in the universe assembled. Correcting for this bias affects published merger rates, galaxy evolution models, and our understanding of how galaxies grow.

The timing of this project matters. The James Webb Space Telescope is finding galaxies in the early universe that are far more massive than theoretical models predict should exist at that age. One leading explanation in the field is that mergers in the early universe have been systematically undercounted. This project demonstrates that such a blind spot exists in the nearby universe and provides a pipeline to address it. The same approach, training a classifier on simulations and deploying it on real survey data, scales directly to JWST imaging.

Looking further ahead, a space-based gravitational wave detector planned for the 2030s will listen for gravitational waves from merging supermassive black holes. Every galaxy merger eventually brings two black holes together. Accurate galaxy merger rates are needed to predict how many events that detector will hear. Better merger catalogs built today directly inform the predictions we make for the next generation of observatories.

The hidden population is real, its physical cause is identified, and a pipeline to find it has been demonstrated to work.

The immediate next step is applying this pipeline to James Webb Space Telescope imaging, where the hidden merger problem is likely more severe. I also plan to pursue spectroscopic follow-up on the hidden merger candidates to confirm the discoveries through kinematic measurements, which would move them from candidates to confirmed detections. Retraining the classifier on high-redshift simulations would extend the approach to the early universe. Extending the deployment to larger surveys with significantly more galaxies would substantially increase the candidate sample. Adding colour information directly to the feature set would better capture the red sequence signature of the hidden population.

I would like to thank Scott Wilkinson, a PhD candidate at the University of Victoria, for providing access to his IllustrisTNG simulation dataset and synthetic galaxy images, partly used to train the classifier in this project. Having access to a properly prepared simulation dataset with known merger histories was essential to the methodology, and I am grateful for his generosity in sharing it.

I would also like to thank the IllustrisTNG collaboration for making their simulation data publicly available; the Galaxy Zoo team and the hundreds of thousands of volunteers whose classifications made the independent validation possible; and the Sloan Digital Sky Survey for over two decades of publicly accessible telescope data.

All research questions, feature engineering, code, analysis, and figures in this project are my own independent work.

Conselice, C. J. (2003). The relationship between stellar light distributions of galaxies and their formation histories. The Astrophysical Journal Supplement Series, 147(1), 1.

Duan, Q., et al. (2025). Galaxy merger rates at high redshift from JWST imaging. Monthly Notices of the Royal Astronomical Society.

Lintott, C. J., et al. (2008). Galaxy Zoo: Morphologies derived from visual inspection of galaxies from the Sloan Digital Sky Survey. Monthly Notices of the Royal Astronomical Society, 389(5), 1179.

Lotz, J. M., Primack, J., & Madau, P. (2004). A new nonparametric approach to galaxy morphological classification. The Astronomical Journal, 128(1), 163.

Nelson, D., et al. (2019). The IllustrisTNG simulations: Public data release. Computational Astrophysics and Cosmology, 6(2).

Rodriguez-Gomez, V., et al. (2019). The optical morphologies of galaxies in the IllustrisTNG simulation. Monthly Notices of the Royal Astronomical Society, 483(4), 4140.

Taylor, E. N., et al. (2011). Galaxy and Mass Assembly: Stellar mass estimates. Monthly Notices of the Royal Astronomical Society, 418(3), 1587.

Wilkinson, S. J., et al. (2024). The limitations and potential of non-parametric morphology statistics for post-merger identification. arXiv:2401.13654.

Wen, Z. Z., & Zheng, X. Z. (2016). Outer asymmetry as a new non-parametric indicator for galaxy morphology. The Astrophysical Journal, 832(1), 90.

Gendler, R. (2008). The Antennae Galaxies (NGC 4038 and NGC 4039) [Photograph]. ESA/Hubble. https://esahubble.org/images/heic0812c/

JWSTFeed. (2025). JWST deep field of Abell 370 [Photograph]. https://jwstfeed.com

Finkelstein, S., Bagley, M., & Levay, Z. (2023). JWST Cosmic Evolution Early Release Science Survey (CEERS) [Photograph]. NASA/STScI. https://ceers.github.io

NASA/CXC/SAO; NASA/ESA/CSA/STScI/Webb. (2025). Composite X-ray and infrared image [Photograph]. Image processing: NASA/CXC/SAO/L. Frattare. https://webbtelescope.org

European Space Agency. (2024). LISA laser interferometer space antenna concept [Illustration]. ESA. https://www.esa.int/Science_Exploration/Space_Science/LISA

National Aeronautics and Space Administration. (2023). JWST deep field high-redshift galaxy candidates [Photograph]. NASA. https://webbtelescope.org

Subaru Telescope, National Astronomical Observatory of Japan. (2024). Hyper Suprime-Cam wide survey field [Photograph]. https://subarutelescope.org