An artificial-intelligence system has produced a list of 29,605 possible pairs of galactic nuclei from images previously rejected by the GOTHIC pipeline. The authors stress that this is a statistically refined candidate catalogue, not a set of confirmed systems.
The preprint asks whether a supervised YOLOv11 detector can sort genuine dual-nucleus candidates from foreground stars and other spurious alignments in the 46,061 galaxies reported as rejected by GOTHIC. The distinction is important because an apparent second object may be an unrelated alignment rather than a second nucleus.
The paper gives two counts for the rejected or testing population: 46,061 in its abstract and research framing, but 46,021 in the testing-set listing. The starting sample size is therefore not stated consistently.
A capable detector, on a limited benchmark
The final detector used YOLOv11x with oriented bounding boxes—boxes that can rotate to match an object’s orientation—and two classes: dual nuclei and foreground stars. Training ran for 237 epochs before early stopping; the selected checkpoint was epoch 137, and the confidence threshold was 0.55.
On held-out labelled validation data, it reported precision of 0.919, recall of 0.905 and an F1 score of 0.912 for the dual-nuclei class at that threshold. Class-specific results showed a different balance: dual-nucleus precision was 0.967 and recall 0.803, with F1 0.877; foreground-star precision was 0.835, recall 0.838 and F1 0.837.
The independent testing set was reserved for post-training inference and candidate discovery, not standard quantitative performance evaluation. That means the validation scores cannot be treated as measured precision or purity for the 29,605-candidate catalogue.
What a human check found
The labelled dataset contained 3,119 images: 1,338 galaxy-merger images, 1,741 star images and 40 bootstrapped mergers. The listed split was 2,495 training images and 624 validation images. Within that development work, the authors manually annotated 1,393 merger candidates, excluded 55 poor-quality or unresolved images, and randomly divided the remaining 1,338 images in an 80:20 split.
For a check on what the output contained, researchers manually inspected a random sample of 200 detections from a 29,460-detection catalogue with valid separation measurements. They judged about 54.5% to be plausible dual-nucleus systems; the other main categories were galaxy-star associations at 24%, single-galaxy detections at 18% and unidentified compact sources at 3.5%.
An analysis of different separation cuts gave estimated bona fide fractions of 54.5% for the full sample, 59% for candidates within 6.87 arcseconds and 62% for those within 9.79 arcseconds. The 95% confidence intervals overlapped, and the differences were not statistically significant at that level.
Applying those sampled fractions to the corresponding valid-detection populations produced statistical projections of about 16,056 systems for the full sample, 13,672 for the 6.87-arcsecond subset and 17,666 for the 9.79-arcsecond subset. The authors present these as projections, not definitive classifications.
Cross-calibration against GOTHIC centroid measurements produced a compact subset defined by a 6.87-arcsecond cut, containing about 13,672 candidates; calibrated separations reached roughly 0.56 arcseconds. The authors describe this as a conservative candidate estimate, not a confirmed census.
The spectra did not settle the question
The study also examined spectroscopy for 183 systems in the sub-kiloparsec subgroup. Forty spectra were too noisy; of the remaining 143 systems, 136 showed no double-peaked emission. Seven were initially flagged, but only one clearly had emission lines, and it was still single-peaked.
Those spectra do not settle whether the compact candidates are genuine pairs: single-fibre SDSS spectroscopy cannot reliably confirm or exclude unresolved sub-kiloparsec pairs. The absence of a double peak is not, by itself, a confirmation or a definitive rejection.
A tool for triage, not a final census
Taken together, the paper offers a way to refine a large rejected image catalogue, but it does not establish that the 29,605 candidates are dual active nuclei. The test output was used for discovery rather than a standard quantitative score, and the manual composition came from a sample rather than exhaustive verification.
The codebase, trained models, preprocessing and post-processing tools, and reproduction instructions are publicly available. Supplementary material includes a catalogue of 212 sub-kiloparsec candidates, while full image cutouts must be regenerated from SDSS.
The document is a preprint dated 25 August 2026 and still carries an unresolved acceptance line. For now, the headline result is a machine-assisted shortlist whose candidates require independent checking.
Paper data and sources
Original title: Decoupling candidate dual AGN from chance superpositions in the GOTHIC survey via a deep-learning framework
Authors: Bhavesh Mukheja, Snehanshu Saha, Anwesh Bhattacharya et al.
Journal/Repository: arXiv
Status: Preprint, not yet peer-reviewed
First online: 2026-08-25
DOI: Not available
Original paper · Full text