Pith. sign in

REVIEW 3 major objections 6 minor 34 references

A convolutional neural network can recover two-dimensional velocity-delay maps from reverberation mapping data, even when the light curves are noisy and have gaps.

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

T0 review · deepseek-v4-flash

2026-08-03 13:20 UTC pith:EYLGOE5U

load-bearing objection A useful within-subfield methods paper that shows a CNN ensemble can deconvolve RM data well inside the training distribution, but generalization to real data is not yet demonstrated — and the paper is honest about that. the 3 major comments →

arxiv 2512.24433 v2 pith:EYLGOE5U submitted 2025-12-30 astro-ph.GA

A Convolutional Neural Network for the Recovery of Transfer Functions From Velocity-Resolved Reverberation Mapping Data

Kirk Long , Keith Horne , Jason Dexter , Benoit Tremblay This is my paper
classification astro-ph.GA
keywords reverberation mappingactive galactic nucleibroad-line regiontransfer functionvelocity-delay mapsconvolutional neural networkdeconvolutionmachine learning
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The paper establishes that a purpose-built convolutional neural network, trained on synthetic reverberation mapping data, can invert the convolution linking AGN continuum variations to emission-line variations and recover the underlying transfer function in both one dimension and full two-dimensional velocity-delay space. The method stays accurate when up to 10–50% of the data are missing, and an ensemble of networks provides realistic confidence intervals without extra computation. Transfer learning lets the same trained ensemble adapt to a new continuum light curve with only a few epochs of fine-tuning. A blind test on more realistic simulated spectra recovers the peak response and red-blue asymmetry, though it leaves a spurious prompt-response artifact that the authors trace to continuum contamination absent from the training set. The result matters because such deconvolved velocity-delay maps are the cleanest observable probe of broad-line region structure around supermassive black holes.

Core claim

The central claim is that the reverberation-mapping deconvolution problem—recovering the response function Ψ(ν, τ) from noisy, gappy line and continuum light curves—can be solved by a convolutional network that learns the operator f(ΔC) such that Ψ = f(ΔC)*ΔL. Trained on tens of thousands of synthetic examples built from damped random walk continua and a library of ring, Gaussian, exponential, disk, cloud, and disk-wind response shapes, the network generalizes to unseen transfer functions, degrades gracefully when observations are missing, and produces uncertainty estimates from an ensemble. The authors demonstrate 1D and 2D recoveries, transfer learning to a different continuum, and agreeme

What carries the argument

The central object is the (D)CNN architecture, a convolutional network with three parallel branches of small, medium, and large temporal and velocity filters, followed by a skip-connection chain and a 1×1 convolution to output the map. It is built on the identity that the inverse operator in the convolution theorem can be learned as a data-driven function f(ΔC) which, convolved with the line light curve, yields the transfer function. Training uses a synthetic data pipeline: a damped-random-walk continuum, transfer functions drawn from three families of shapes plus additive combinations, 0.5% white noise, and randomly dropped data points. The ensemble of ~20–40 networks supplies empirical err

Load-bearing premise

The method's validity rests on the assumption that synthetic training data—generated by convolving a damped random walk continuum with a library of hand-built transfer functions, adding 0.5% white noise and random gaps, and assuming perfectly de-trended, continuum-free light curves—are representative enough of real reverberation mapping data that a network trained on them will recover true transfer functions from observations.

What would settle it

Feed the trained ensemble a set of light curves from a real AGN whose velocity-delay response has been independently mapped by a fully physical model (e.g., one with sharp edges and a red-blue asymmetry), and check whether the ensemble's 1σ recovery region contains that independent map. If it does not, or if the spurious prompt-response artifact seen in the paper's blind test persists after continuum contamination is added to the training set as the authors propose, the method's generalization claim would be falsified.

Watch this falsifier. Get emailed when new claim-graph text bears on it.

If this is right

  • If correct, the method offers a cheap, fast alternative to regularized inversion codes for producing velocity-delay maps from current and future 2D reverberation campaigns, with uncertainties computed directly from the ensemble.
  • The method could be applied to any reverberation deconvolution problem, including accretion disk and torus reverberation mapping, not just the broad-line region.
  • Because transfer learning requires only a few epochs to adapt to a new continuum, the same trained ensemble can be fine-tuned for individual sources, easing the analysis burden expected from upcoming survey data.
  • The blind test demonstrates that the recovered maps agree with existing analytic inversions, suggesting the method can be used as a cross-check on those results.
  • The failure mode identified in the blind test—spurious prompt response due to continuum contamination—points to a concrete improvement: adding continuum contamination to the training set would likely resolve it.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • Editorial inference: The authors' own interpretability tests suggest the multi-scale split architecture may be unnecessary; a simpler skip-only network might perform equally well and be faster to train, a direction they explicitly flag for future work.
  • Editorial inference: Because the method depends on the assumed linear convolution model, it will likely inherit systematics from nonlinear or time-varying response behavior; a testable extension is probing recovery on simulations with nonlinear line response.
  • Editorial inference: Since the training set excludes continuum contamination, real-data applications with imperfect continuum subtraction will tend to produce spurious prompt response (as in the blind test); a practical workaround is co-training with contaminated light curves, which the paper suggests for the future.
  • Editorial inference: The ensemble uncertainty estimates appear calibrated on synthetic data, but a stronger reliability test would apply the method to a source with an independent geometric or dynamical model and compare the recovered map directly.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit.

Referee Report

3 major / 6 minor

Summary. This manuscript presents a convolutional neural network ensemble for recovering the one- and two-dimensional transfer functions (velocity-delay maps) from synthetic reverberation-mapping light curves. The network is trained on mock line light curves produced by convolving a single DRW continuum with a mixture of analytic basis shapes, an accretion-disk prescription, and kinematic BLR models, with white noise and randomized gaps. On held-out synthetic data from the same generator the method recovers 1D and 2D transfer functions accurately, degrades gracefully with 10–50% missing data, produces ensemble uncertainties that visually track residuals, and compares favorably with MEMEcho. Transfer learning to a different continuum is demonstrated, and a blind test on Mangham et al. synthetic spectra reproduces the main velocity-delay structure but adds a spurious prompt-response component, which the authors attribute to continuum contamination absent from training. Interpretability experiments show that the physically motivated scale-split filters do not activate in the expected way.

Significance. The in-distribution results are credible, and the combination of a simple CNN architecture with an ensemble for uncertainty estimation is a useful practical contribution for RM analysis. The reproducibility assets (code and sample models) are a positive feature. However, the broad claim that the method 'can successfully deconvolve reverberation mapping data products' is not yet established for real or out-of-distribution data, and the key conceptual limitation that the continuum is not an input to the network needs to be addressed or prominently disclosed. If these points are fixed, the paper would be a solid methods contribution to the RM community.

major comments (3)
  1. [§2.1–2.2, Eq. (1)–(3)] The network input is the line light curve/spectrogram alone; the continuum is used only to generate the training data. Thus the learned map is not the general inverse f(ΔC) in Eq. (3), but a supervised regression from ΔL to Ψ that is conditional on the specific DRW continuum used in training and on the hand-built Ψ families. The paper acknowledges this in §5, but the abstract and §7 statements that the method can 'successfully deconvolve reverberation mapping data products' and could be applied to 'any reverberation deconvolution problem' overstate the current evidence. Please qualify these claims, e.g., by describing the method as a continuum-conditioned supervised inversion rather than a general deconvolution.
  2. [§5, Fig. 7] The only out-of-family blind test shows a systematic artifact: extra prompt response at τ≈0 that is not present in the MEMEcho recovery. The authors' explanation (continuum contamination absent from the training set) is plausible, but it means the method has not been validated on the type of input it will encounter in real RM data. All quantitative successes in Figs. 2–6 are generated by the same pipeline used for training. I recommend adding a test in which continuum contamination is included in the training set, or carrying out a more aggressive continuum-subtraction step and showing that the artifact is removed, or explicitly restricting the central claim to in-distribution synthetic data.
  3. [§3.1, §4, Fig. 5d] The claims that 'the error bar always encapsulates the true response function' and that confidence intervals are 'sensible' are based on visual inspection of in-distribution validation examples. No coverage fraction is reported, and the blind test in Fig. 7 does not compare the predicted uncertainty map to the true residual. Because the ensemble spread accounts for training and input noise but not systematic errors from distribution shift, the uncertainty should be presented as conditional on the training distribution unless a quantitative coverage test on out-of-distribution data is added.
minor comments (6)
  1. [§2.2/References] Horne et al. 1991a and 1991b are listed as separate references but appear to be the same paper (ApJL 367, L5). Please correct the citation.
  2. [§2.2] All training transfer functions are normalized, so the method predicts only the shape of Ψ, not its absolute responsivity. This is an important limitation and should be stated in the abstract or introduction.
  3. [§5] The test is called 'blind' even though the Mangham et al. answer is public. Consider using 'out-of-family simulated test' to avoid confusion.
  4. [Figure 10] The architecture diagram is dense; adding the tensor shape (nT, nLC, nC) at each stage would improve reproducibility.
  5. [§3.1, Fig. 3 right] The MEMEcho comparison uses different missing-data masks for the line and continuum light curves. Please clarify whether this asymmetry affects the comparison.
  6. [§8] The availability statement should include a URL or persistent repository link for the code and models.

Circularity Check

0 steps flagged

No significant circularity: validation targets are withheld during training and the blind test provides independent out-of-family grounding.

full rationale

The paper's central claim is that a (D)CNN can deconvolve reverberation mapping data products. The method is trained on synthetic lightcurves generated by convolving a DRW continuum with hand-built transfer-function families, then evaluated on a 30% validation split. The validation ground truths are not used in gradient descent—the paper states the model never sees this data in each gradient descent update—so the reported recoveries are genuine predictions rather than fitted values. The main quantitative demonstrations are in-distribution with respect to the training generator, which limits external validity but is not circularity: the input-to-target map is learned, not encoded by definition. The out-of-family Mangham blind test uses synthetic spectra from a different simulation approach and is not part of the training set; its imperfect recovery, including the spurious prompt response, is explicitly acknowledged and attributed to continuum contamination absent from training, and the paper even reports that the test inputs lie outside the training lightcurve distribution. This is an honest external benchmark, not a self-referential validation. The self-citations to Long et al. 2023 and Long & Dexter 2025 supply kinematic BLR model shapes used only as training-set generators; they do not justify the deconvolution claim, are not invoked as a uniqueness theorem, and do not force the CNN predictions. The term "(D)CNN" is a descriptive label for a convolutional network trained on a deconvolution task, not a renamed known result. The remaining weaknesses—limited training diversity, single-continuum fitting in the core tests, and no out-of-distribution coverage test for the confidence intervals—are correctness and generalization concerns, not circular reasoning.

Axiom & Free-Parameter Ledger

8 free parameters · 6 axioms · 0 invented entities

The method contributes an inversion prior learned from simulated data; it introduces no new physics or entities. The cost is a large set of hand-chosen hyperparameters and domain assumptions about the realism of synthetic training data. The most consequential assumption is that de-trended, continuum-subtracted, DRW-driven light curves with white noise cover the real data distribution; the blind test shows this assumption is already violated.

free parameters (8)
  • CNN filter size scales = kts=10, ktm=20, ktl=200 for nT=1000; kvs=(2,2), kvm=(5,5), kvl=(7,10) for 25/50 velocity bins
    Chosen arbitrarily to represent small/medium/large scales; §2.1.
  • Learning rate = ~1e-4
    Found experimentally; §2.2.
  • Noise level and dropout fraction = 5e-3 dynamic range; 10% random gaps
    Mimic first-order observational errors; §2.2.
  • Ensemble size and pruning threshold = ~100 models; keep within 1σ of validation loss
    Arbitrary choice; §2.2.
  • Training mixture weights = basis+disk 1/16 each, BLR/disk-wind 1/8 each, combinations 1/2
    Hand-set to balance shape and physical models; §2.2.
  • Early stopping window = 5 epochs
    Chosen experimentally; §2.2.
  • Maximum training delay = ~100 days (baseline/10)
    Imposed so model only learns delays observable in 1000-day campaign; §3.
  • Loss function = mean square error
    Standard choice; alternatives deferred to future work.
axioms (6)
  • domain assumption Line variations are a linear, time-invariant convolution of continuum variations (Eq. 1).
    Foundation of RM; used to generate all training/test data; deviations are acknowledged as possible in §1.
  • domain assumption Training transfer functions drawn from hand-built families (basis shapes, face-on disk, kinematic BLR) are representative of real BLR responses.
    The CNN learns the inversion prior from these families; §2.2; the blind test partly challenges this.
  • domain assumption Continuum variability can be modeled by a damped random walk and noise as white Gaussian at 0.5% dynamic range.
    Authors state DRW is an oversimplification; used for training data; §2.2.
  • domain assumption Input lightcurves are perfectly de-trended and continuum-subtracted.
    Training set has no continuum contamination; §2.2. The blind test's prompt-response artifact is attributed to violation of this premise in §5.
  • ad hoc to paper Only normalized transfer-function shape is targeted, not absolute intensity.
    All training Ψ are normalized; §2.2. This limits the claim to shape recovery.
  • ad hoc to paper Validation performance on synthetic data generated with the same pipeline is a meaningful measure of inversion accuracy.
    Held-out sets share the forward model with training; §3, §4; the external Mangham test is the only check outside this.

pith-pipeline@v1.3.0-alltime-deepseek · 17461 in / 13321 out tokens · 131166 ms · 2026-08-03T13:20:44.987558+00:00 · methodology

0 comments
read the original abstract

One of the hallmarks of active galactic nuclei are that they are highly variable with time. In watching the spectra vary it has been observed that the emission-lines often appear to "reverberate" -- that is they vary in response to continuum variations assumed to originate close to the black hole. This critical observation underlies the reverberation mapping technique, an elegant physics experiment that has allowed us to characterize the environment around many supermassive black holes in nearby active galactic nuclei. Recent observations are of such quality that the response can be measured as a function of velocity across the emission-line, and in doing so we can construct velocity-delay maps that show the structure and physics of the gas in the broad-line region better than any other measurement to date. Unfortunately constructing such maps requires a deconvolution, and given that the data are often noisy and with gaps such deconvolutions are non-trivial. Here we present a novel deconvolution method for the recovery of velocity-delay maps using a custom convolutional neural network architecture, showcasing that such methods have great promise for the deconvolution of reverberation mapping data products. While we have designed this new method with the BLR in mind, in principle this technique could be applied to any reverberation deconvolution problem, including in the accretion disk and torus.

Figures

Figures reproduced from arXiv: 2512.24433 by Benoit Tremblay, Jason Dexter, Keith Horne, Kirk Long.

Figure 1
Figure 1. Figure 1: Top: Sample 1D transfer functions for each “ba￾sis” function used in training the model. Middle: Sample lightcurves resulting from the convolution of the correspond￾ing transfer function from the top panel with the continuum. Bottom: The DRW continuum lightcurve convoled with the top panel to generate the middle panel. signal likely does not extend to the BLR, we in￾clude this as it is another nice represe… view at source ↗
Figure 2
Figure 2. Figure 2: Top grid: Sample recoveries of each of the six “basis” transfer functions shown in [PITH_FULL_IMAGE:figures/full_fig_p006_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Comparing how the novel (D)CNN and existing analytic MEMEcho techniques perform when the training data quality are degraded Left: Recoveries of a more complicated 1D transfer function that is the result of the combination of several of the “basis” functions shown in [PITH_FULL_IMAGE:figures/full_fig_p007_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: The distribution of errors for the sample ensemble of 2D models. Full examples of each model transfer function and lightcurve inputs are in [PITH_FULL_IMAGE:figures/full_fig_p007_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: a: Sample full 2D transfer function of a complicated combined synthetic BLR with both a “blob” component superimposed on a more traditional virial “disk-wind” component. b: (D)CNN model recovery of the full 2D transfer function. c: Residuals (prediction in top middle panel - ground truth shown in top left panel). Regions where the model overpredicts are shown in red, underpredicted regions are shown in blu… view at source ↗
Figure 6
Figure 6. Figure 6: A 2D inversion of another complicated combined transfer function, this time the result of a two “shape” BLRs (an exponential decay component + a gaussian blob). Panels a-g are the same as described in the caption of [PITH_FULL_IMAGE:figures/full_fig_p010_6.png] view at source ↗
Figure 7
Figure 7. Figure 7: (D)CNN inversion of lightcurves extracted from a simulated quasar BLR dataset first presented in S. W. Mangham et al. (2019). a: mean predicted Ψ(λ, t). b: uncertainty on mean prediction. c: 1D transfer function Ψ(τ ) with error bar corresponding to ±1σ uncertainty. d: 1D Ψ(λ) with error bar corresponding to ±1σ uncertainty. e: mean loss curves for the ensemble of models [PITH_FULL_IMAGE:figures/full_fig_… view at source ↗
Figure 8
Figure 8. Figure 8: Similarity distributions of the lightcurves used in training/validating the model, with the vertical line indicat￾ing how the set of extracted lightcurves for the quasar BLR in S. W. Mangham et al. (2019) compares to the distribu￾tion used in training. “black-box” than one might like. Interestingly the only trend between the two input lightcurves is the change in importance of the “skip” and “norm/dropout”… view at source ↗
Figure 9
Figure 9. Figure 9: Here we showcase the percentage of activations (how strongly the model responds to an input and thus how much each feature contributes to the final prediction) for two very different input lightcurves. Top: The input lightcurves are shown in blue and red, while the continuum is shown for reference in black. Middle: The transfer functions used to generate the lightcurves in the top panel by convolving with … view at source ↗
Figure 10
Figure 10. Figure 10: A schematic of the general architecture of our (D)CNN. The indigo boxes represent the input, output, and convolutional layers, with each convolutional layer including a description of the number of channels at that step (we fix nC = 1 in this work). The green boxes represent regularization—BatchNorm (BN) and Dropout (Drop)—and activation—ReLU—layers that occur after each convolutional layer. The red diamo… view at source ↗
Figure 11
Figure 11. Figure 11: Similar to [PITH_FULL_IMAGE:figures/full_fig_p017_11.png] view at source ↗

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

34 extracted references · 4 canonical work pages

  1. [1]

    2019, arXiv e-prints, arXiv:1902.05569, doi: 10.48550/arXiv.1902.05569

    Akeson, R., Armus, L., Bachelet, E., et al. 2019, arXiv e-prints, arXiv:1902.05569, doi: 10.48550/arXiv.1902.05569

    Pith/arXiv arXiv doi:10.48550/arxiv.1902.05569 2019

  2. [2]

    2022, Frontiers in Astronomy and Space Sciences, 9, doi: 10.3389/fspas.2022.1001043

    Michalewicz, K. 2022, Frontiers in Astronomy and Space Sciences, 9, doi: 10.3389/fspas.2022.1001043

    arXiv 2022

  3. [4]

    D., & McKee, C

    Blandford, R. D., & McKee, C. F. 1982b, ApJ, 255, 419, doi: 10.1086/159843

    doi:10.1086/159843

  4. [5]

    M., Bentz, M

    Cackett, E. M., Bentz, M. C., & Kara, E. 2021, iScience, 24, 102557, doi: 10.1016/j.isci.2021.102557

    arXiv 2021

  5. [6]

    2025, The Astrophysical Journal, 980, 257, doi: 10.3847/1538-4357/adae85

    Deesamutara, S., Chainakun, P., Worrakitpoonpon, T., et al. 2025, The Astrophysical Journal, 980, 257, doi: 10.3847/1538-4357/adae85

    doi:10.3847/1538-4357/adae85 2025

  6. [7]

    2010, in Proceedings of Machine Learning Research, Vol

    Glorot, X., & Bengio, Y. 2010, in Proceedings of Machine Learning Research, Vol. 9, Proceedings of the Thirteenth International Conference on Artificial Intelligence and Statistics, ed. Y. W. Teh & M. Titterington (Chia Laguna Resort, Sardinia, Italy: PMLR), 249–256. https://proceedings.mlr.press/v9/glorot10a.html GRA VITY Collaboration, Abuter, R., Accar...

    Pith/arXiv arXiv doi:10.48550/arxiv.2301.08071 2010

  7. [8]

    P., Bechtol, K., Bellm, E., et al

    Guy, L. P., Bechtol, K., Bellm, E., et al. 2024, Rubin Observatory Plans for an Early Science Program, doi: 10.5281/zenodo.11110648

    doi:10.5281/zenodo.11110648 2024

  8. [9]

    2015, CoRR, abs/1512.03385

    He, K., Zhang, X., Ren, S., & Sun, J. 2015, CoRR, abs/1512.03385

    Pith/arXiv arXiv 2015

  9. [10]

    2018, Journal of Cosmology and Astroparticle Physics, 2018, 054, doi: 10.1088/1475-7516/2018/07/054

    Lucchi, A. 2018, Journal of Cosmology and Astroparticle Physics, 2018, 054, doi: 10.1088/1475-7516/2018/07/054

    doi:10.1088/1475-7516/2018/07/054 2018

  10. [11]

    2023, ApJ, 948, 85, doi: 10.3847/1538-4357/acc45a

    Homayouni, Y., De Rosa, G., Plesha, R., et al. 2023, ApJ, 948, 85, doi: 10.3847/1538-4357/acc45a

    doi:10.3847/1538-4357/acc45a 2023

  11. [12]

    1994, in Astronomical Society of the Pacific Conference Series, Vol

    Horne, K. 1994, in Astronomical Society of the Pacific Conference Series, Vol. 69, Reverberation Mapping of the Broad-Line Region in Active Galactic Nuclei, ed. P. M

    1994

  12. [14]

    F., & Peterson, B

    Horne, K., Welsh, W. F., & Peterson, B. M. 1991b, The Astrophysical Journal Letters, 367, L5, doi: 10.1086/185919

    doi:10.1086/185919

  13. [15]

    M., et al

    Horne, K., De Rosa, G., Peterson, B. M., et al. 2021, ApJ, 907, 76, doi: 10.3847/1538-4357/abce60

    doi:10.3847/1538-4357/abce60 2021

  14. [16]

    2018, Journal of Open Source Software, doi: 10.21105/joss.00602

    Innes, M. 2018, Journal of Open Source Software, doi: 10.21105/joss.00602

    doi:10.21105/joss.00602 2018

  15. [17]

    2018, CoRR, abs/1811.01457

    Innes, M., Saba, E., Fischer, K., et al. 2018, CoRR, abs/1811.01457. https://arxiv.org/abs/1811.01457

    Pith/arXiv arXiv 2018

  16. [18]

    P., Vogeley, M

    Kasliwal, V. P., Vogeley, M. S., & Richards, G. T. 2015, MNRAS, 451, 4328, doi: 10.1093/mnras/stv1230

    doi:10.1093/mnras/stv1230 2015

  17. [19]

    T., Alloin, D., Barr, P., et al

    Korista, K. T., Alloin, D., Barr, P., et al. 1995, ApJS, 97, 285, doi: 10.1086/192144 ML inversions of RM measurements15 Koz lowski, S., Kochanek, C. S., Udalski, A., et al. 2010, ApJ, 708, 927, doi: 10.1088/0004-637X/708/2/927

    work page doi:10.1086/192144 1995

  18. [20]

    A., De Rosa, G., Ely, J., et al

    Kriss, G. A., De Rosa, G., Ely, J., et al. 2019, ApJ, 881, 153, doi: 10.3847/1538-4357/ab3049

    doi:10.3847/1538-4357/ab3049 2019

  19. [21]

    H., & Done, C

    Krolik, J. H., & Done, C. 1995, ApJ, 440, 166, doi: 10.1086/175258

    doi:10.1086/175258 1995

  20. [22]

    H., Horne, K., Kallman, T

    Krolik, J. H., Horne, K., Kallman, T. R., et al. 1991, ApJ, 371, 541, doi: 10.1086/169918

    doi:10.1086/169918 1991

  21. [23]

    L., Yuan, C., Bruschweiler-Li, L., & Br¨ uschweiler, R

    Li, D.-W., Hansen, A. L., Yuan, C., Bruschweiler-Li, L., & Br¨ uschweiler, R. 2021, Nature Communications, 12, doi: 10.1038/s41467-021-25496-5

    doi:10.1038/s41467-021-25496-5 2021

  22. [24]

    2022, GEOPHYSICS, 87, S249–S265, doi: 10.1190/geo2020-0904.1

    Liu, C., Sun, M., Dai, N., et al. 2022, GEOPHYSICS, 87, S249–S265, doi: 10.1190/geo2020-0904.1

    work page doi:10.1190/geo2020-0904.1 2022

  23. [25]

    2025, The Astrophysical Journal, 987, 196, doi: 10.3847/1538-4357/adda38

    Long, K., & Dexter, J. 2025, The Astrophysical Journal, 987, 196, doi: 10.3847/1538-4357/adda38

    work page doi:10.3847/1538-4357/adda38 2025

  24. [26]

    2023, The Astrophysical Journal, 953, 184, doi: 10.3847/1538-4357/ace4bb

    Long, K., Dexter, J., Cao, Y., et al. 2023, The Astrophysical Journal, 953, 184, doi: 10.3847/1538-4357/ace4bb

    work page doi:10.3847/1538-4357/ace4bb 2023

  25. [27]

    L., Ivezi´ c,ˇZ., Kochanek, C

    MacLeod, C. L., Ivezi´ c,ˇZ., Kochanek, C. S., et al. 2010, ApJ, 721, 1014, doi: 10.1088/0004-637X/721/2/1014

    doi:10.1088/0004-637x/721/2/1014 2010

  26. [28]

    W., Knigge, C., Williams, P., et al

    Mangham, S. W., Knigge, C., Williams, P., et al. 2019, MNRAS, 488, 2780, doi: 10.1093/mnras/stz1713

    doi:10.1093/mnras/stz1713 2019

  27. [29]

    Peterson, B. M. 1993, PASP, 105, 247, doi: 10.1086/133140

    doi:10.1086/133140 1993

  28. [30]

    J., et al

    Wang, S., Woo, J.-H., Barth, A. J., et al. 2025, The Astrophysical Journal, 983, 45, doi: 10.3847/1538-4357/adbca5

    doi:10.3847/1538-4357/adbca5 2025

  29. [31]

    D., Vestergaard, M., & Davis, T

    Watson, D., Denney, K. D., Vestergaard, M., & Davis, T. M. 2011, ApJL, 740, L49, doi: 10.1088/2041-8205/740/2/L49

    doi:10.1088/2041-8205/740/2/l49 2011

  30. [32]

    F., & Horne, K

    Welsh, W. F., & Horne, K. 1991, ApJ, 379, 586, doi: 10.1086/170530

    doi:10.1086/170530 1991

  31. [33]

    R., Pancoast, A., Treu, T., et al

    Williams, P. R., Pancoast, A., Treu, T., et al. 2020, ApJ, 902, 74, doi: 10.3847/1538-4357/abbad7

    doi:10.3847/1538-4357/abbad7 2020

  32. [34]

    S., Liu, C., & Jia, J

    Xu, L., Ren, J. S., Liu, C., & Jia, J. 2014, in Advances in Neural Information Processing Systems, ed. Z. Ghahramani, M. Welling, C. Cortes, N. Lawrence, & K. Weinberger, Vol. 27 (Curran Associates, Inc.). https://proceedings.neurips.cc/paper files/paper/2014/ file/1c1d4df596d01da60385f0bb17a4a9e0-Paper.pdf

    2014

  33. [35]

    Graham, M. J. 2022, The Astrophysical Journal, 936, 132, doi: 10.3847/1538-4357/ac8351

    doi:10.3847/1538-4357/ac8351 2022

  34. [36]

    S., et al

    Zhang, Z.-X., Du, P., Smith, P. S., et al. 2019, ApJ, 876, 49, doi: 10.3847/1538-4357/ab1099 16Long et al. APPENDIX A.(D)CNN MODEL ARCHITECTURE SUMMARY Input(nT, nLC, nC ) BN+ReLU Split Conv2D(kts, 1); nC → 32 BN+ReLU+Drop Conv2D(1, kvs); 32 → 32 BN+ReLU+Drop Conv2D(ktm, 1); nC → 32 BN+ReLU+Drop Conv2D(1, kvm); 32 → 32 BN+ReLU+Drop Conv2D(ktl, 1); nC → 32...

    doi:10.3847/1538-4357/ab1099 2019