pith. sign in

arxiv: 2606.04491 · v1 · pith:Z2LH3NDFnew · submitted 2026-06-03 · ⚛️ nucl-th

NuGNN: a Graph Neural Network for Nuclear Reaction Network Equations

C. H. Kim , K. Y. Chae , S. Ko , M. R. Mumpower , M. S. Smith This is my paper

Pith reviewed 2026-06-28 04:26 UTC · model grok-4.3

classification ⚛️ nucl-th
keywords graph neural networksnuclear reaction networkssurrogate modelingastrophysical simulationsX-ray burstsisotope abundancesdeep learning
0
0 comments X

The pith

A graph neural network surrogate for nuclear reaction networks reproduces final isotope abundance patterns when substituted for the original solver.

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

Nuclear reaction networks for astrophysical simulations are slowed by stiffness and repeated calls to Jacobian-based implicit solvers, especially for networks with hundreds of isotopes. The paper introduces NuGNN, a graph neural network that encodes the reaction network directly through separate isotope and reaction nodes with message passing along the actual reaction links. Trained on data covering wide ranges of temperature, density, and time-step size, NuGNN delivers errors of only a few percent. When inserted into the network evolution code, it alone reproduces the final abundance patterns of the original solver, while fully connected networks and Res-U-Nets do not. The model also reduces computation time, showing a route to faster large-scale simulations.

Core claim

NuGNN directly reflects the structure of the reaction network through heterogeneous isotope and reaction nodes and message-passing along reaction connections. When implemented in the network evolution code in place of the original solver, NuGNN successfully reproduces the final abundance patterns, whereas the other architectures fail to do so. The trained model can substantially improve computational speed for a 690-isotope network under general Type I X-ray burst conditions.

What carries the argument

Heterogeneous graph with isotope nodes, reaction nodes, and message-passing along reaction connections used as a surrogate solver for the stiff network equations.

If this is right

  • NuGNN achieves significantly better accuracy than Res-U-Net or fully connected networks, with errors of only a few percent.
  • The model can be substituted directly into existing network evolution codes and still recover correct final abundances.
  • The trained surrogate substantially reduces computation time compared with the original implicit solver.
  • The graph architecture succeeds on a 690-isotope network under conditions spanning many orders of magnitude in temperature and density.

Where Pith is reading between the lines

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

  • The same graph construction could be applied to other large reaction networks whose topology is known, potentially extending the approach beyond X-ray bursts.
  • Because the model respects the reaction graph, retraining on data from different astrophysical regimes might require less data than dense networks.
  • Success here suggests that preserving the explicit reaction connectivity in the architecture is the key property that allows the surrogate to remain stable when integrated over many time steps.

Load-bearing premise

The training data spanning many orders of magnitude in temperature, density, and time-step size is sufficient to cover all conditions encountered during actual network evolution under Type I X-ray burst conditions.

What would settle it

A full network evolution run with NuGNN in place of the solver that produces final abundance patterns differing by more than a few percent from the original solver under X-ray burst conditions.

Figures

Figures reproduced from arXiv: 2606.04491 by C. H. Kim, K. Y. Chae, M. R. Mumpower, M. S. Smith, S. Ko.

Figure 1
Figure 1. Figure 1: Schematic input–output overview of NuGNN for a single abundance update. NuGNN processes the input quantities using a graph neural network composed of isotope and reaction nodes and predicts the corresponding change in mass fraction, ∆X. See the text for details [PITH_FULL_IMAGE:figures/full_fig_p003_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Transformed labels under different prepro￾cessing schemes, shown as functions of ∆X (∆X/X). sign(∆X) · log10(|∆X|) is referenced to the upper x-axis (∆X), whereas the others are referenced to the lower x-axis (∆X/X). ∆X/X spans a narrower dynamic range than ∆X. sign(∆X) · (C + log10(|∆X/X|)) provides a more compact range while preserving the correct magnitude ordering and avoiding a disconnection at zero. … view at source ↗
Figure 3
Figure 3. Figure 3: Overview of NuGNN architecture. (a) Illustration of the graph composed of two different types of nodes: isotope and reaction nodes. (b) Examples of message-passing where each node exchanges messages with neighboring nodes connected through nuclear reactions. (c) Full model architecture, including flux preprocessing, embedding, message-passing blocks, and the two-channel sigmoid output head. See the text fo… view at source ↗
Figure 4
Figure 4. Figure 4: Learning curves for NuGNN, Res-U-Net, and FNN. The validation loss indicates that NuGNN outperforms the other two architectures. For clarity, only the epoch with the lowest RMS within each 5-epoch interval is shown. style additions between hidden layers. The final output layer follows the same two-channel method described in Section 4.1.6. It consists of five fully connected layers in total, with a hidden … view at source ↗
Figure 5
Figure 5. Figure 5: Examples of NuGNN predictions for the test dataset. NuGNN reproduces the true ∆X (Label) remarkably well, even though the data span tens of orders of magnitude in both the negative and positive directions [PITH_FULL_IMAGE:figures/full_fig_p009_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: Final abundances obtained with different solvers for different thermodynamic trajectories from J. Fisker et al. (2005); J. L. Fisker et al. (2008) which were not used for training. ‘Original’ denotes the final abundances evolved using the original solver. The title of each subplot indicates the zone number of the simulations of J. Fisker et al. (2005); J. L. Fisker et al. (2008). While Res-U-Net and FNN we… view at source ↗
Figure 7
Figure 7. Figure 7: shows the computing times we measured on test data samples for different solvers. We used In￾tel Xeon Platinum 8360Y for the original solver and NVIDIA RTX 5090 for the surrogates. The inference times of the surrogates do not depend on the input data as shown in the figure. Graph neural networks are gen￾erally slower than conventional neural networks because they rely on computationally expensive operation… view at source ↗
read the original abstract

Nuclear reaction networks are a major computational bottleneck in astrophysical simulations when large isotope sets are required, because of the stiffness of the network equations and the repeated calls to Jacobian-based solvers required by implicit methods. In this work, we develop a deep learning surrogate solver for a large 690-isotope nuclear reaction network under general Type I X-ray burst conditions using a graph neural network, NuGNN. Unlike conventional fully connected or convolutional neural networks, NuGNN directly reflects the structure of the reaction network through heterogeneous isotope and reaction nodes and message-passing along reaction connections. The model is trained on data spanning many orders of magnitude in stellar temperature and density and in simulation time step size. We compare NuGNN with a Res-U-Net and fully connected neural network and find that NuGNN consistently achieves significantly better accuracy with errors of only a few percent. More importantly, when implemented in the network evolution code in place of the original solver, NuGNN successfully reproduces the final abundance patterns, whereas the other architectures fail to do so. We also show that the trained model can substantially improve computational speed, demonstrating its practical potential for large-scale simulations. These results show that graph neural networks provide a robust and promising framework for accurate surrogate modeling of large nuclear reaction networks.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Referee Report

2 major / 1 minor

Summary. The manuscript introduces NuGNN, a graph neural network surrogate for solving the stiff equations of a 690-isotope nuclear reaction network under Type I X-ray burst conditions. The architecture uses heterogeneous isotope and reaction nodes with message passing along reaction edges. Trained on data spanning wide ranges of temperature, density, and time-step size, NuGNN is reported to achieve few-percent errors and to outperform both a Res-U-Net and a fully connected network. The central result is that, when substituted for the implicit solver inside the network evolution code, NuGNN reproduces the final abundance patterns while the other architectures do not, and it also provides a computational speedup.

Significance. If the closed-loop substitution result holds under varied conditions and time-stepping logic, the work would demonstrate a practical route to accelerating large nuclear networks in astrophysical simulations by replacing repeated Jacobian-based implicit solves with a fast, structure-preserving surrogate. The explicit encoding of the reaction graph is a methodological strength that distinguishes the approach from generic neural-network surrogates.

major comments (2)
  1. [Abstract] Abstract: the claim that NuGNN 'successfully reproduces the final abundance patterns' when implemented in the network evolution code is load-bearing for the practical utility argument, yet no quantitative metrics are supplied on per-isotope or integrated abundance errors over the full sequence of time steps, nor on whether the test used the same adaptive time-stepping logic as the production run.
  2. [Methods] Methods (training-data generation): the statement that training data span 'many orders of magnitude in stellar temperature and density and in simulation time step size' does not specify whether samples were drawn independently or along full evolutionary trajectories, nor whether the distribution covers the correlated abundance vectors and time-step sequences actually encountered during closed-loop integration; this directly affects the coverage assumption required for stable long-term evolution.
minor comments (1)
  1. The phrase 'errors of only a few percent' should be accompanied by the precise error metric (e.g., mean relative error per species, maximum error, or L2 norm) and by the distribution of errors across isotopes.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for their constructive review and for recognizing the potential significance of encoding the reaction network structure in NuGNN. We respond point-by-point to the major comments below.

read point-by-point responses

  1. Referee: [Abstract] Abstract: the claim that NuGNN 'successfully reproduces the final abundance patterns' when implemented in the network evolution code is load-bearing for the practical utility argument, yet no quantitative metrics are supplied on per-isotope or integrated abundance errors over the full sequence of time steps, nor on whether the test used the same adaptive time-stepping logic as the production run.

    Authors: We agree that quantitative metrics would strengthen the closed-loop result. In the revised manuscript we will add per-isotope relative errors and integrated abundance discrepancies evaluated over the full sequence of time steps in the substitution test, together with an explicit statement confirming that the test employed the identical adaptive time-stepping logic used in the production runs. revision: yes

  2. Referee: [Methods] Methods (training-data generation): the statement that training data span 'many orders of magnitude in stellar temperature and density and in simulation time step size' does not specify whether samples were drawn independently or along full evolutionary trajectories, nor whether the distribution covers the correlated abundance vectors and time-step sequences actually encountered during closed-loop integration; this directly affects the coverage assumption required for stable long-term evolution.

    Authors: The training data were generated by sampling along full evolutionary trajectories under the relevant conditions rather than from independent snapshots, precisely to ensure coverage of the correlated abundance vectors and time-step sequences that appear during closed-loop integration. We will expand the Methods section with an explicit description of this trajectory-based sampling procedure and the resulting coverage of the relevant parameter space. revision: yes

Circularity Check

0 steps flagged

No circularity; surrogate trained and tested on independent external data

full rationale

The paper trains NuGNN on simulation data generated from the original solver across wide ranges of T, rho, and dt, then substitutes the model into the independent network evolution code to check reproduction of final abundances. No equations reduce to fitted quantities by construction, no self-citation chains support load-bearing uniqueness claims, and no ansatzes or renamings are smuggled in. The derivation chain remains self-contained against external benchmarks.

Axiom & Free-Parameter Ledger

1 free parameters · 1 axioms · 0 invented entities

The central claim rests on the learned mapping from initial conditions to evolved abundances; the graph architecture encodes the reaction topology but the weights are fitted to simulation output. No new physical entities are introduced.

free parameters (1)
  • Neural network weights and biases
    All model parameters are fitted during supervised training on simulation trajectories.
axioms (1)
  • domain assumption The nuclear reaction network can be faithfully represented as a static heterogeneous graph whose edges capture all relevant reaction channels.
    Invoked by the choice of graph construction and message-passing scheme.

pith-pipeline@v0.9.1-grok · 5774 in / 1276 out tokens · 35100 ms · 2026-06-28T04:26:31.583560+00:00 · methodology

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

43 extracted references · 34 canonical work pages · 2 internal anchors

  1. [1]

    W., Beers, T

    Arcones, A., Bardayan, D. W., Beers, T. C., et al. 2017, Progress in Particle and Nuclear Physics, 94, 1, doi: https://doi.org/10.1016/j.ppnp.2016.12.003

    work page doi:10.1016/j.ppnp.2016.12.003 2017

  2. [2]

    Boehnlein, M

    Boehnlein, A., Diefenthaler, M., Sato, N., et al. 2022, Rev. Mod. Phys., 94, 031003, doi: 10.1103/RevModPhys.94.031003

    work page doi:10.1103/revmodphys.94.031003 2022

  3. [3]

    2020, Monthly Notices of the Royal Astronomical Society, 494, 3037, doi: 10.1093/mnras/staa910

    Bravo, E. 2020, Monthly Notices of the Royal Astronomical Society, 494, 3037, doi: 10.1093/mnras/staa910

    work page doi:10.1093/mnras/staa910 2020

  4. [4]

    W., Blondin, J

    Bruenn, S. W., Blondin, J. M., Hix, W. R., et al. 2020, The Astrophysical Journal Supplement Series, 248, 11, doi: 10.3847/1538-4365/ab7aff

    work page doi:10.3847/1538-4365/ab7a 2020

  5. [5]

    J., Sneden, C., Lawler, J

    Cowan, J. J., Sneden, C., Lawler, J. E., et al. 2021, Reviews of Modern Physics, 93, 015002, doi: 10.1103/RevModPhys.93.015002

    work page doi:10.1103/revmodphys.93.015002 2021

  6. [6]

    2022, The Astrophysical Journal, 940, 134, doi: 10.3847/1538-4357/ac9a4b

    Nonaka, A. 2022, The Astrophysical Journal, 940, 134, doi: 10.3847/1538-4357/ac9a4b

    work page doi:10.3847/1538-4357/ac9a4b 2022

  7. [7]

    E., Petermann, I., et al

    Farmer, R., Fields, C. E., Petermann, I., et al. 2016, The Astrophysical Journal Supplement Series, 227, 22, doi: 10.3847/1538-4365/227/2/22

    work page doi:10.3847/1538-4365/227/2/22 2016

  8. [8]

    2005, Nuclear Physics A, 752, 604, doi: https://doi.org/10.1016/j.nuclphysa.2005.02.063

    Fisker, J., Brown, E., Liebend¨ orfer, M., Thielemann, F.-K ., & Wiescher, M. 2005, Nuclear Physics A, 752, 604, doi: https://doi.org/10.1016/j.nuclphysa.2005.02.063

    work page doi:10.1016/j.nuclphysa.2005.02.063 2005

  9. [9]

    L., Schatz, H., & Thielemann, F.-K

    Fisker, J. L., Schatz, H., & Thielemann, F.-K. 2008, The Astrophysical Journal Supplement Series, 174, 261, doi: 10.1086/521104

    work page doi:10.1086/521104 2008

  10. [10]

    2016, in Proceedings of Machine Learning Research, Vol

    Gal, Y., & Ghahramani, Z. 2016, in Proceedings of Machine Learning Research, Vol. 48, Proceedings of The 33rd International Conference on Machine Learning, ed. M. F. Balcan & K. Q. Weinberger (New York, New York, USA: PMLR), 1050–1059 13

    2016

  11. [11]

    Dahl, G. E. 2017, in Proceedings of Machine Learning

    2017

  12. [12]

    2015, in Proceedings of the IEEE International Conference on Computer Vision (ICCV) (Santiago, Chile: IEEE Computer Society), 1440–1448

    Girshick, R. 2015, in Proceedings of the IEEE International Conference on Computer Vision (ICCV) (Santiago, Chile: IEEE Computer Society), 1440–1448

    2015

  13. [13]

    , keywords =

    Grichener, A., Renzo, M., Kerzendorf, W. E., et al. 2025, The Astrophysical Journal Supplement Series, 279, 49, doi: 10.3847/1538-4365/ade717

    work page doi:10.3847/1538-4365/ade717 2025

  14. [14]

    2016, in Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR) (Las Vegas, NV, USA: IEEE Computer Society)

    He, K., Zhang, X., Ren, S., & Sun, J. 2016, in Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR) (Las Vegas, NV, USA: IEEE Computer Society)

    2016

  15. [15]

    R., & Meyer, B

    Hix, W. R., & Meyer, B. S. 2006, Nuclear Physics A, 777, 188, doi: https://doi.org/10.1016/j.nuclphysa.2004.10.009 Ioffe, S., & Szegedy, C. 2015, in Proceedings of the 32nd International Conference on Machine Learning, Vol. 37 (Lille, France: PMLR), 448–456

    work page doi:10.1016/j.nuclphysa.2004.10.009 2006

  16. [16]

    , keywords =

    Jermyn, A. S., Bauer, E. B., Schwab, J., et al. 2023, The Astrophysical Journal Supplement Series, 265, 15, doi: 10.3847/1538-4365/acae8d

    work page doi:10.3847/1538-4365/acae8d 2023

  17. [17]

    2022, The Astrophysical Journal, 929, 96, doi: 10.3847/1538-4357/ac5f09

    Kim, C., Chae, K., Cha, S., et al. 2022, The Astrophysical Journal, 929, 96, doi: 10.3847/1538-4357/ac5f09

    work page doi:10.3847/1538-4357/ac5f09 2022

  18. [18]

    H., Chae, K

    Kim, C. H., Chae, K. Y., & Smith, M. S. 2026, Phys. Rev. C, 113, 024308, doi: 10.1103/mcxf-d32x

    work page doi:10.1103/mcxf-d32x 2026

  19. [19]

    H., Chae, K

    Kim, C. H., Chae, K. Y., Smith, M. S., et al. 2024, Phys. Rev. C, 110, 054609, doi: 10.1103/PhysRevC.110.054609

    work page doi:10.1103/physrevc.110.054609 2024

  20. [20]

    P., & Ba, J

    Kingma, D. P., & Ba, J. 2015, in Proceedings of the 3rd International Conference on Learning Representations (ICLR), San Diego, CA, USA

    2015

  21. [21]

    2017, in Advances in Neural Information Processing Systems, ed

    Lakshminarayanan, B., Pritzel, A., & Blundell, C. 2017, in Advances in Neural Information Processing Systems, ed. I. Guyon, U. V. Luxburg, S. Bengio, H. Wallach, R. Fergus, S. Vishwanathan, & R. Garnett, Vol. 30 (Red

    2017

  22. [22]

    Lampe, N., Heger, A., & Galloway, D. K. 2016, The Astrophysical Journal, 819, 46, doi: 10.3847/0004-637X/819/1/46

    work page doi:10.3847/0004-637x/819/1/46 2016

  23. [23]

    Lippuner, J., & Roberts, L. F. 2017, The Astrophysical Journal Supplement Series, 233, 18, doi: 10.3847/1538-4365/aa94cb

    work page doi:10.3847/1538-4365/aa94cb 2017

  24. [24]

    L., Hannun, A

    Maas, A. L., Hannun, A. Y., & Ng, A. Y. 2013, in Proceedings of the ICML Workshop on Deep Learning for Audio, Speech, and Language Processing, Atlanta,

    2013

  25. [25]

    2018, ApJ, 860, 147, doi: https://doi.org/10.3847/1538-4357/aac3d3 M¨ oller, P., Sierk, A

    Meisel, Z. 2018, ApJ, 860, 147, doi: https://doi.org/10.3847/1538-4357/aac3d3 M¨ oller, P., Sierk, A. J., Ichikawa, T., & Sagawa, H. 2016, Atomic Data and Nuclear Data Tables, 109, 1, doi: 10.1016/j.adt.2015.10.002

    work page doi:10.3847/1538-4357/aac3d3 2018

  26. [26]

    R., Kawano, T., Korobkin, O., Misch, G

    Mumpower, M. R., Kawano, T., Korobkin, O., Misch, G. W., & Sprouse, T. M. 2025, Atomic Data and Nuclear Data Tables, 165, 101736, doi: 10.1016/j.adt.2025.101736 Nav´ o, G., Reichert, M., Obergaulinger, M., & Arcones, A. 2023, The Astrophysical Journal, 951, 112, doi: 10.3847/1538-4357/acd640

    work page doi:10.1016/j.adt.2025.101736 2025

  27. [27]

    2019, in Advances in Neural Information Processing Systems, Vol

    Paszke, A., Gross, S., Massa, F., et al. 2019, in Advances in Neural Information Processing Systems, Vol. 32 (Red

    2019

  28. [28]

    2015, ApJS, 220, 15, doi: 10.1088/0067-0049/220/1/15

    Paxton, B., Marchant, P., Schwab, J., et al. 2015, ApJS, 220, 15, doi: https://doi.org/10.1088/0067-0049/220/1/15

    work page internal anchor Pith review doi:10.1088/0067-0049/220/1/15 2015

  29. [29]

    In: Proc

    Courville, A. 2018, Proceedings of the AAAI Conference on Artificial Intelligence, 32, doi: 10.1609/aaai.v32i1.11671

    work page doi:10.1609/aaai.v32i1.11671 2018

  30. [30]

    Gated Attention for Large Language Models: Non-linearity, Sparsity, and Attention-Sink-Free

    Qiu, Z., Wang, Z., Zheng, B., et al. 2025, arXiv e-prints, arXiv:2505.06708, doi: 10.48550/arXiv.2505.06708

    work page internal anchor Pith review Pith/arXiv arXiv doi:10.48550/arxiv.2505.06708 2025

  31. [31]

    2023, The Astrophysical Journal Supplement Series, 268, 66, doi: 10.3847/1538-4365/acf033

    Reichert, M., Winteler, C., Korobkin, O., et al. 2023, The Astrophysical Journal Supplement Series, 268, 66, doi: 10.3847/1538-4365/acf033

    work page doi:10.3847/1538-4365/acf033 2023

  32. [32]

    2015, in Medical Image Computing and Computer-Assisted Intervention – MICCAI 2015, ed

    Ronneberger, O., Fischer, P., & Brox, T. 2015, in Medical Image Computing and Computer-Assisted Intervention – MICCAI 2015, ed. N. Navab, J. Hornegger, W. M. Wells, & A. F. Frangi (Cham: Springer International Publishing), 234–241

    2015

  33. [33]

    The graph neural network model.IEEE Transactions on Neural Networks, 20(1):61–80, 2009

    Monfardini, G. 2009, IEEE Transactions on Neural Networks, 20, 61, doi: 10.1109/TNN.2008.2005605

    work page doi:10.1109/tnn.2008.2005605 2009

  34. [34]

    2006, Nuclear Physics A, 777, 601, doi: https://doi.org/10.1016/j.nuclphysa.2005.05.200

    Schatz, H., & Rehm, K. 2006, Nuclear Physics A, 777, 601, doi: https://doi.org/10.1016/j.nuclphysa.2005.05.200

    work page doi:10.1016/j.nuclphysa.2005.05.200 2006

  35. [35]

    2001, NuPhA, 688, 150, doi: 10.1016/S0375-9474(01)00688-1

    Schatz, H., Aprahamian, A., Barnard, V., et al. 2001, NuPhA, 688, 150, doi: 10.1016/S0375-9474(01)00688-1

    work page doi:10.1016/s0375-9474(01)00688-1 2001

  36. [36]

    S., & Lu, D

    Smith, M. S., & Lu, D. 2024, Frontiers in Astronomy and Space Sciences, Volume 11 - 2024, doi: 10.3389/fspas.2024.1494439

    work page doi:10.3389/fspas.2024.1494439 2024

  37. [37]

    M., Mumpower, M

    Sprouse, T. M., Mumpower, M. R., & Surman, R. 2021, Phys. Rev. C, 104, 015803, doi: 10.1103/PhysRevC.104.015803

    work page doi:10.1103/physrevc.104.015803 2021

  38. [38]

    Timmes, F. X. 1999, The Astrophysical Journal Supplement Series, 124, 241, doi: 10.1086/313257

    work page doi:10.1086/313257 1999

  39. [39]

    X., Hoffman, R

    Timmes, F. X., Hoffman, R. D., & Woosley, S. E. 2000, The Astrophysical Journal Supplement Series, 129, 377, doi: 10.1086/313407

    work page doi:10.1086/313407 2000

  40. [40]

    2013, Frontiers of Physics, 8, 199, doi: 10.1007/s11467-013-0315-y 14 Veliˇ ckovi´ c, P., Cucurull, G., Casanova, A., et al

    Travaglio, C., & Raphael Hix, W. 2013, Frontiers of Physics, 8, 199, doi: 10.1007/s11467-013-0315-y 14 Veliˇ ckovi´ c, P., Cucurull, G., Casanova, A., et al. 2018, in Proceedings of the 6th International Conference on Learning Representations (ICLR), Vancouver, BC, Canada

    work page doi:10.1007/s11467-013-0315-y 2013

  41. [41]

    Wang, X., He, X., Wang, M., Feng, F., & Chua, T.-S. 2019, in Proceedings of the 42nd International ACM SIGIR Conference on Research and Development in Information Retrieval (New York, NY, USA: Association for Computing Machinery), 165–174, doi: 10.1145/3331184.3331267

    work page doi:10.1145/3331184.3331267 2019

  42. [42]

    2025, The Astrophysical Journal, 990, 105, doi: 10.3847/1538-4357/adf331

    Zhang, X., Yi, Y., Wang, L., et al. 2025, The Astrophysical Journal, 990, 105, doi: 10.3847/1538-4357/adf331

    work page doi:10.3847/1538-4357/adf331 2025

  43. [43]

    L., Lund, K

    Zhu, Y. L., Lund, K. A., Barnes, J., et al. 2021, ApJ, 906, 94, doi: 10.3847/1538-4357/abc69e

    work page doi:10.3847/1538-4357/abc69e 2021