PaperlayerView exact revision on arXiv

Source-linked AI summary

Decoupling candidate dual AGN from chance superpositions in the GOTHIC survey via a deep-learning framework

Bhavesh Mukheja, Snehanshu Saha, Anwesh Bhattacharya, Mousumi Das, Françoise Combes, Sudhanshu Barway

arXiv:2608.24164v1astro-ph.GAcs.CVcs.LGmath-phphysics.data-an

TL;DR

Dual AGN are difficult to identify because compact nuclei can be confused with foreground stars, unresolved structure, and chance projections. The paper applies supervised YOLOv11 oriented-box detection to 46,061 rejected GOTHIC candidates, producing 29,605 dual-nucleus candidates and a conservative compact subset of ∼13,672, while emphasizing that confirmation requires higher-resolution follow-up.

  • Problem

    Compact dual nuclei are difficult to distinguish from foreground stars, unresolved substructure, and chance projections in large imaging surveys.

  • Method

    The study trains a supervised YOLOv11 oriented-bounding-box detector on annotated SDSS imaging to separate dual nuclei from stellar contaminants.

  • Results

    29,605 dual-nucleus candidates were identified, with ∼54.5–62% estimated as genuine dual-nucleus systems and ∼13,672 retained in the compact d≤6.87'' regime.

  • Takeaways & Limitations

    Deep learning refines the plausible dual-nucleus census, including compact candidates that earlier selection rejected or could not resolve cleanly.

  • Takeaways & Limitations

    Most candidates lack spectroscopic redshifts for both nuclei, so physical association and exclusion of chance projections remain unconfirmed.

Abstract

from arXiv · show

Dual active galactic nuclei (DAGN) mark a critical phase in the evolution of merging galaxies and the pairing of supermassive black holes, yet they remain difficult to identify in large imaging surveys because of projection effects and limited spatial resolution. Compact foreground stars and unresolved substructure can mimic dual nuclei through chance superposition, complicating automated detection. We revisit the 46,061 galaxies flagged but rejected as DAGN candidates by the GOTHIC pipeline, primarily because the two nuclei fell within the SDSS fibre aperture or exceeded its separation threshold. We train a supervised deep-learning framework based on the YOLOv11 oriented-bounding-box architecture on annotated SDSS imaging to separate genuine dual nuclei from foreground stellar contaminants and other spurious alignments. The final model attains a validation precision of 0.919, recall of 0.905, and $F_1$ of 0.912 for the dual-nuclei class, and yields 29,605 dual-nucleus candidates after removing star-dominated and blended detections. Structured visual inspection indicates that $54.5$--$62\%$ are consistent with genuine dual nuclei, implying $\sim(1.4$--$1.8)\times10^{4}$ plausible systems. Cross-calibrating the YOLO separation against the deterministic GOTHIC centroid measurement and restricting to the compact regime ($d \le 6.87''$) gives a conservative subset of $\sim 13{,}672$ candidates, reaching calibrated separations of $\sim 0.56''$. Spectroscopy of the most compact ($\le 1$~kpc) systems shows they are dominated by passive, absorption-line galaxies with no resolved double-peaked emission, so confirmation requires higher-resolution follow-up. The catalogue is a statistically refined list of candidates, not confirmed DAGN. Nonetheless, deep-learning detection substantially reduces contamination and expands the plausible DAGN census.

1 INTRODUCTION

Galaxy mergers can bring supermassive black holes into paired systems, making dual AGN valuable probes, but compact nuclei are difficult to distinguish from projections and contaminants. This study applies YOLOv11-OBB to rejected GOTHIC candidates to refine the candidate census.

  • Dual AGN provide electromagnetic probes of SMBH pairing, merger-driven black-hole growth, and the final stages of galaxy mergers.
  • Confirmed DAGN are rare, with reported dual-AGN fractions of ∼1.3% in optical SDSS pairs and ∼7.5% in hard X-ray-selected pairs.
  • Double-peaked emission lines and imaging-based pair selection provide useful candidates but suffer contamination, projection effects, and unresolved-substructure ambiguities.
  • Large-survey methods remain biased against compact systems and constrained by SDSS fibre collisions and limited spectroscopic resolution.
  • The study trains a supervised YOLOv11-OBB detector on 46,061 rejected GOTHIC galaxies to distinguish dual nuclei from foreground stars and identify candidates rather than confirmed DAGN.
  • The model yields 29,605 candidates, with structured inspection estimating ∼54.5–62% as genuine dual-nucleus systems and ∼13,672 in the compact d≤6.87'' subset.

2 PRELIMINARIES

YOLO is a single-stage detector that jointly localizes and classifies sources, making it suitable for crowded astronomical images. Here, an oriented YOLOv11 workflow is applied to annotated SDSS cutouts and a heterogeneous, spectroscopically constrained GOTHIC-derived sample.

  • YOLO formulates object detection as one-stage regression, predicting bounding boxes, confidence scores, and class probabilities in a single CNN pass.
  • Object detection adds source localization to image classification, which is essential for crowded, overlapping, or structurally complex astronomical images.
  • The pipeline predicts class labels, confidence scores, and oriented-box coordinates, with IoU used to evaluate predicted-region overlap.
  • YOLOv11n serves as a lightweight baseline, while YOLOv11x-OBB provides rotation-aware detection because projected DAGN morphologies have arbitrary sky orientations.
  • The workflow comprises annotation, quality control, an 80:20 training–validation split, model monitoring, and application to an independent test dataset.
  • The GOTHIC-derived sample was conservatively rejected because the 3'' fibre can blend nuclei, while blending, misassociation, and unresolved substructure make the population heterogeneous.

3 METHODOLOGY

The methodology develops and evaluates a rotation-aware YOLOv11x-OBB detector trained on annotated SDSS imaging to distinguish dual nuclei from stellar and spurious detections.

  • Dataset and pipeline: The workflow uses annotated SDSS cutouts for training, validation, inference, and post-processing into a dual-nuclei catalogue.The final pipeline includes supervised training, large-scale inference, and filtering of detections.
  • Model development: The staged development strategy begins with a minimal baseline to diagnose morphology-driven limitations before adopting a more expressive rotation-aware detector.The baseline exposed contamination, poor handling of inclined structures, and duplicate detections, motivating the final architecture.
  • Catalogue construction: Post-processing reduces the catalogue to 29,605 dual-nucleus candidates from 46,021 test objects, while manual inspection identifies foreground stars, faint objects, and single-nucleus misclassifications.The final model reduces the relative frequency of these failure modes compared with earlier analyses.
  • Model development: The detector assigns dual galactic nuclei and foreground stars to separate classes, while spatial relationships define overlapping, exclusive, and other detection categories.This classification supports post-processing of ambiguous or blended sources.
  • Final model evaluation: At confidence threshold 0.55, the final model achieves validation precision 0.919, recall 0.905, and F1 score 0.912 for the dual-nuclei class.The selected checkpoint is epoch 137, supported by peak validation mAP and F1 with stable training and validation curves.
  • Final model evaluation: Validation diagnostics show strong class separation, with approximately 97% precision and 80% recall for dual-nuclei detections in the class-wise evaluation.These class-wise figures are distinct from the aggregate model-level metrics reported at the selected operating threshold.

4 RESULTS

The final YOLOv11x-OBB pipeline removes foreground-star contaminants and refines rejected GOTHIC detections into a statistically characterized dual-nucleus candidate catalogue. Calibration and manual inspection indicate substantial sensitivity to compact systems, but residual ambiguity and SDSS resolution limit definitive classification.

  • Post-processing and false-positive mitigation: 33,190 raw detections were produced from 46,021 SDSS objects before post-processing and categorization.Detections were assigned to dual galactic nuclei and foreground-star classes, with spatial relationships defining the filtering categories.
  • Post-processing and false-positive mitigation: 29,605 dual-nucleus candidates remained after filtering, a 36% reduction driven by removing foreground-star contaminants.This refined sample is smaller than both the 31,250-candidate baseline YOLO output and the original 46,021-object GOTHIC output.
  • Residual detections and limitations: Residual ambiguity persists because surface-brightness sensitivity, limited higher-redshift resolution, projection effects, and catalogue incompleteness cannot be fully resolved from single-band SDSS morphology.The retained detections therefore include physically distinct scenarios despite reduced failure-mode frequency in the rotation-aware two-class model.
  • Manual inspection and reliability assessment: 54.5% of the full sample, 59% of the d≤6.87′′ subset, and 62% of the d≤9.79′′ subset were estimated to be bona fide dual-nucleus systems.The overlapping confidence intervals indicate that purity differences across subsets are not statistically significant at the 95% confidence level.
  • Projected separation estimation: ∼0.56′′ was the calibrated centroid-to-centroid separation corresponding to the compact subset, although this value is an extrapolation with largest relative scatter for the closest pairs.The calibration gives GOTHIC centroid separations averaging 3.1′′ below YOLO OBB edges and supports sensitivity to closely separated components.
  • Model performance: 0.919 precision, 0.905 recall, and 0.912 F1 were achieved for the dual-nuclei class on the validation dataset.These labelled-data metrics do not directly provide the absolute purity of the unlabelled catalogue.
  • Projected candidate counts: ∼13,672 likely dual-nucleus candidates were adopted as a conservative estimate within the d≤6.87′′ regime.Applying the estimated fractions yielded projected counts of ∼16,056 for the full sample, ∼13,672 for d≤6.87′′, and ∼17,666 for d≤9.79′′.
  • Interpretation: The catalogue expands the candidate census by suppressing foreground-star contamination while retaining sensitivity to compact nuclear systems.The present candidates are statistically inferred and are not directly comparable to individually confirmed GOTHIC nuclei pairs or spectroscopically established dual AGN.

Redshift reliability

Reliable spectroscopic redshifts are required to convert angular separations into physical units, and very small redshifts can make compactness indistinguishable from measurement error.

  • Redshift reliability: 29 of 212 systems had z<10^-4, below the general SDSS redshift uncertainty, so their physical separations could not be reliably interpreted.These systems were excluded, leaving 183 for spectroscopic analysis.

Method and diagnostic

Single-fibre spectroscopy provides a diagnostic for sub-kiloparsec candidates, but instrumental resolution, projection, blended spectra, and impure line signatures constrain confirmation. The examined systems were predominantly passive or lacked resolved double-peaked emission, so higher-resolution follow-up remains necessary.

  • Spectroscopic diagnostic: 183 systems were examined after excluding 29 with unreliable redshifts, and each sub-kiloparsec candidate produced one spatially blended SDSS-fibre spectrum.A resolved double-peaked narrow emission line would be the canonical unresolved dual-AGN diagnostic.
  • Physical caveats: R∼2000 corresponds to a velocity FWHM of ∼150 km s^-1, limiting whether line-of-sight velocity splitting can be resolved.Random orbital-plane projection can further reduce the observable velocity component, so a missing double peak does not exclude a genuine dual nucleus.
  • Physical caveats: Double-peaked narrow lines are not unique to a second nucleus because rotating gas discs, AGN outflows, and jet–ISM interactions can produce them.A positive double-peak detection is therefore only a candidate signature rather than confirmation.
  • Spectroscopic results: 136 of 143 analysable spectra showed no indication of double-peaked emission, while 40 of 183 systems were too noisy for meaningful line analysis.The analysable spectra were dominated by Balmer absorption and near-featureless continua, consistent with predominantly passive stellar populations.
  • Interpretation and follow-up: No confirmed double-peaked-emission dual-nucleus candidate was found within the ≤1 kpc subset.Single-fibre blending, SDSS resolution, projection, and the impurity of double-peaked lines explain why this negative result does not by itself establish absence of dual nuclei.

5 DISCUSSION

The study uses supervised deep learning to recover compact dual-nucleus candidates rejected by GOTHIC, while quantifying contamination and emphasizing that statistical candidates require follow-up confirmation.

  • Motivation and comparison: 46,061 galaxies rejected by GOTHIC were reassessed with a supervised deep-learning framework designed to distinguish nuclear structures from stellar contaminants.The framework was motivated by failures of purely image-processing methods on compact, ambiguous, or low-contrast structures.
  • Model performance: 0.919 precision and 0.905 recall were achieved for the dual-nucleus class on validation data, while the full rejected sample yielded 29,605 candidates.The model was iteratively refined to address confusion between compact galactic nuclei and stars.
  • Contamination and inspection: 54.5% of 200 inspected detections were plausible dual-nucleus systems, with galaxy–star associations comprising 24% and single-galaxy false positives 18%.The misclassification proportions were broadly consistent across sampled compact-separation subsets.
  • Limitations: Limited training diversity for close galaxy–star superpositions leaves residual ambiguity, especially when stellar profiles resemble compact nuclei.Such cases are underrepresented in the annotated training set, particularly at close angular proximity and low signal-to-noise.
  • Contamination and inspection: Approximately 3.5% of detections were unidentified compact sources, potentially reflecting catalog incompleteness, faint companions, or difficult imaging conditions.These sources lacked clear counterparts in SDSS photometric or spectroscopic catalogs.
  • Limitations: Spectroscopic redshifts for both nuclei are generally unavailable, so projected separations rely on assuming both components share the primary galaxy’s redshift.The ≤1 kpc systems are dominated by passive absorption-line spectra, and the lone clear emission-line case is single-peaked at SDSS resolution.

6 CONCLUSION

The YOLO-based framework refines a previously rejected SDSS sample into a much larger set of dual-nucleus candidates, but the resulting catalogue remains statistically inferred rather than spectroscopically confirmed.

  • Conclusion: 46,061 previously excluded galaxies were analyzed with YOLO-based detection using raw SDSS imaging, a dedicated star class, and oriented bounding boxes.The design explicitly addresses foreground stellar contamination and rotational variance.
  • Conclusion: 0.919 precision, 0.905 recall, and 0.912 F1-score were obtained on the validation dataset.The final model was a YOLOv11x-OBB configuration whose learned morphology distinctions suppress a dominant source of false positives.
  • Conclusion: 29,605 candidates were produced, with manual inspection indicating 54.5%–62% plausible dual-nucleus systems and approximately 1.4 × 10^4–1.8 × 10^4 likely genuine configurations.The compact d≤6.87′′ regime contains approximately 13,672 physically relevant candidates.
  • Conclusion: The model identifies separations calibrated to approximately 0.56′′ without spectroscopic preselection, denoising, or handcrafted morphological features.This value corresponds to an approximately 4.61′′ YOLO OBB-edge separation.
  • Conclusion: The catalogue is a statistically refined candidate list rather than a definitive census because projection effects, irregular morphologies, and unresolved sources remain in single-band imaging.Confirmation requires spectroscopic or higher-resolution follow-up.

DATA AVAILABILITY

The study provides reproducibility resources while distributing the large SDSS-derived dataset through regeneration instructions rather than hosting all image products directly.

  • Reproducibility: The codebase, trained models, preprocessing utilities, post-processing scripts, and reproduction instructions are publicly available in the project repository.The repository is hosted at https://github.com/BhaveshMukheja/yolo-dagn.
  • Catalogues: A catalogue of 212 sub-kiloparsec candidates and their SDSS cutouts is available as supplementary material.
  • Data access: The imaging data come from SDSS DR16, while scripts and instructions regenerate image cutouts and intermediate products from the original SDSS sources.The full dataset is not hosted in the repository because of storage constraints and survey distribution policies.
  • Catalogues: Additional derived catalogues are available from the corresponding author upon reasonable request.

APPENDIX A: YOLO ARCHITECTURE

YOLO performs joint object localization and classification, while YOLOv11-OBB extends detection to rotated geometries suited to arbitrarily oriented nuclear structures.

  • Architecture: YOLO formulates detection as a single-stage dense prediction problem that jointly optimizes localization and classification in one forward pass.
  • Architecture: The network maps an input image to a multi-scale detection tensor containing bounding-box parameters, objectness scores, and class probabilities.
  • Architecture: The backbone extracts hierarchical feature maps, the neck fuses spatial and semantic information, and the detection head regresses geometry, objectness, and class probabilities.
  • Oriented bounding boxes: YOLOv11-OBB represents objects with four ordered corner points, allowing arbitrary quadrilateral geometries for inclined, elongated, or asymmetric nuclear structures.
  • Optimization: The composite training loss combines oriented-box regression, class-probability, and objectness terms, with parameters optimized by stochastic gradient descent.
  • Training setup: YOLOv11x-OBB was initialized with pretrained DOTA-v1.5 weights, and annotations were generated with CVAT.
  • Evaluation: Validation used an 80:20 annotated split, while the independent test set was reserved solely for post-training inference and candidate discovery.
  • Evaluation: mAP50-95 was used for checkpoint selection and early stopping, alongside precision, recall, F1-score, and related diagnostics.Precision measures the fraction of predicted detections corresponding to genuine nuclei; recall measures recovered ground-truth components; F1 balances both.

APPENDIX D: PROJECTED SEPARATION ESTIMATION ALGORITHM

Projected separations are estimated from YOLO oriented bounding boxes in image space, converted to angular units, then mapped to physical distances using the primary nucleus redshift.

  • Separation objective: The algorithm targets projected physical separation because three-dimensional separations are unavailable for most detections.It combines image-plane geometry with available spectroscopic information.
  • Image-plane measurement: The YOLO model encloses both nuclei in an oriented bounding box represented by four normalized image-space corners.Coordinates are converted to pixels for a fixed 120 × 120-pixel image.
  • Image-plane measurement: The maximum of the four OBB edge lengths provides a conservative upper bound on projected nuclear separation.This ensures both components are enclosed within the estimated distance.
  • Angular conversion: The angular separation is obtained by multiplying the pixel estimate by the SDSS scale of 0.3 arcsec pixel−1.
  • Physical conversion: Physical separation is calculated as angular separation in radians times the angular diameter distance, assuming both nuclei share the primary nucleus redshift.The calculation uses a low-redshift linear Hubble-law approximation.
  • Approximation: The low-redshift distance approximation overestimates exact ΛCDM angular diameter distances by a few percent near z≈0.1 and tens of percent by z≈0.3.Absolute physical separations should therefore be interpreted with this limitation.

APPENDIX E: CONFIDENCE INTERVAL ESTIMATION FOR SENSITIVITY ANALYSIS

The appendix models manual classification as a binomial process and uses confidence intervals to quantify uncertainty in estimated bona fide dual-nucleus fractions and population-level results.

  • Binomial model: Manual classification is represented as X∼Binomial(n,p), where p is the probability that a detection is a genuine dual nucleus.
  • Uncertainty estimation: For sufficiently large n, the estimator’s sampling distribution is approximated as normal, with variance and standard error derived from the binomial model.
  • Confidence intervals: A 95% confidence interval uses the critical value z0.975≈1.96.
  • Sensitivity samples: The compact-separation subset is defined by d≤6.87′′ and has n=100 and k=59, with a 95% confidence interval of [0.494, 0.686].
  • Application: The resulting intervals are used to assess statistical consistency across subsets and propagate uncertainty into population-level estimates.

F3 Errors-in-variables regression

The calibration compares YOLO and GOTHIC separations using errors-in-variables regression and agreement analysis, then adopts an angular Deming relation for operational conversion.

  • Regression choice: Deming regression with λ=1 is adopted because both separation measurements contain error, unlike ordinary least squares.Reduced major axis regression provides a cross-check.
  • Physical calibration: For the physical pair, Pearson r=0.903 and Deming gives E≈0.77C−2.05, with slope bracket [0.72, 0.88].The relation is also described as consistent with a purely multiplicative form.
  • Agreement analysis: Bland–Altman analysis finds GOTHIC averages 3.12′′ below YOLO for angular separations, with 95% limits of [−5.81, −0.44]′′ and proportional bias.The difference shrinks as separation grows, with slope +0.149 and p≈10−6.
  • Agreement analysis: The physical ratio E/C has mean 0.663 with limits [0.32, 1.01] and is essentially flat across the separation range.
  • Adopted calibration: The operational conversion uses the angular Deming relation, after which physical separations apply the same per-object angular-diameter scale k(z).

THE SUB-KILOPARSEC CANDIDATE SAMPLE

The supplement presents 212 sub-kiloparsec YOLO detections as dual-nucleus candidates, while emphasizing that SDSS cannot resolve the two nuclei at these separations.

  • Sample definition: The sample contains 212 YOLOv11x–OBB detections with GOTHIC-calibrated centroid separations sGOTHIC≤1 kpc.These form the leftmost bin of the projected physical-separation distribution.
  • Resolution: At sub-kiloparsec separations, the two nuclei are unresolved in SDSS imaging.
  • Interpretation: The listed objects are dual-nucleus candidates rather than confirmed dual AGN.
Loading 2608.24164v1…