pith. sign in

arxiv: 2606.04991 · v1 · pith:7XSDZP53new · submitted 2026-06-03 · 🌌 astro-ph.SR

Deep Learning with Magnetic Parameter Constraints for Short-Term Prediction of Solar Active Region Vector Magnetic Fields

Yuqing Zhou , Hui Liu , Zhenyu Jin , Yuyang Li , Sizhong Zou , Jiaben Lin , Mingfu Shao , Zhuoheng Huang This is my paper

Pith reviewed 2026-06-28 04:08 UTC · model grok-4.3

classification 🌌 astro-ph.SR
keywords solar active regionsvector magnetic fieldsdeep learningmagnetic constraintsspace weatherSDO/SHARP magnetogramsshort-term predictionunsigned flux
0
0 comments X

The pith

A deep learning model with magnetic parameter constraints predicts solar active region vector magnetic fields 12 hours ahead.

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

The authors build an end-to-end neural network to forecast the full vector magnetic field in solar active regions. They represent the field in three channels, apply dynamic masks to focus training on strong-field patches, and add multiple magnetic-parameter terms to the loss function. On SDO/SHARP data the model reaches SSIM 0.912 and CC 0.998 for the radial component while holding unsigned-flux error to 7.82 percent. These numbers matter because reliable short-term evolution forecasts are needed for space-weather warnings. The work shows that image-level accuracy and physical-quantity consistency can be achieved together in one training loop.

Core claim

The proposed model, trained with dynamic masks of active regions and multi-parameter magnetic constraints, achieves horizon-averaged SSIM of 0.912 and CC of 0.998 for the radial magnetic field component Br, with RMSE between 13 and 21 G. Horizontal components reach SSIM values of 0.728 to 0.800 with CC above 0.895. Unsigned magnetic flux is predicted with an error of 7.82 percent (95 percent CI +/-0.11 percent). This demonstrates both strong performance in image space and consistency with magnetic diagnostics.

What carries the argument

Multi-parameter magnetic constraints added to the training loss, combined with dynamic masks on three-channel vector-magnetogram inputs, to enforce consistency across the 12-hour forecast horizon.

If this is right

  • The radial component Br is forecasted with SSIM above 0.9 and correlation 0.998.
  • Horizontal components maintain SSIM between 0.73 and 0.80 and correlation above 0.89.
  • Unsigned flux errors remain at 7.82 percent with narrow confidence interval.
  • The predictions stay consistent under the magnetic-parameter diagnostics used in the study.
  • The method offers initial support for future space-weather forecasting pipelines.

Where Pith is reading between the lines

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

  • The same constraint strategy could be tested on longer forecast horizons to check whether physical consistency holds beyond 12 hours.
  • Adding further invariants such as force-free conditions might reduce residual inconsistencies in the horizontal components.
  • Real-time integration with operational magnetogram streams would be a direct next step to assess practical utility.
  • Comparison against unconstrained image-prediction baselines would quantify how much the magnetic terms improve flux preservation.

Load-bearing premise

That the added magnetic-parameter terms in the loss will keep the predicted vector fields consistent with observed physical quantities without creating new inconsistencies in the components.

What would settle it

An independent test set where the unsigned magnetic flux error exceeds 8 percent across a statistically significant number of active regions.

Figures

Figures reproduced from arXiv: 2606.04991 by Hui Liu, Jiaben Lin, Mingfu Shao, Sizhong Zou, Yuqing Zhou, Yuyang Li, Zhenyu Jin, Zhuoheng Huang.

Figure 1
Figure 1. Figure 1: Schematic overview of the proposed model architecture for a single time [PITH_FULL_IMAGE:figures/full_fig_p006_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Comparison of ground-truth and predicted magnetic field components for HARP 7959 [PITH_FULL_IMAGE:figures/full_fig_p009_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Same as Figure 2 but for HARP 8026 (NOAA AR 12953, 2022-02-25 10:00 and [PITH_FULL_IMAGE:figures/full_fig_p010_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Pixel-value scatter comparison between predicted and reference fields for HARP 7959 [PITH_FULL_IMAGE:figures/full_fig_p011_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: Same as Figure 4 but for HARP 8026 (NOAA AR 12953). [PITH_FULL_IMAGE:figures/full_fig_p011_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: Br prediction and residual analysis for HARP 7959 (NOAA AR 12934). Left panels: ground truth (top), predicted (middle), and residual (bottom) maps at forecast hours 6 and 12. Right panel: residual distribution histogram aggregated over both time steps, with mean and ±1σ indicated. A diverging red–blue colormap is used for field values and a coolwarm colormap for residuals [PITH_FULL_IMAGE:figures/full_fig… view at source ↗
Figure 7
Figure 7. Figure 7: Same as Figure 6 but for HARP 8026 (NOAA AR 12953). [PITH_FULL_IMAGE:figures/full_fig_p012_7.png] view at source ↗
Figure 8
Figure 8. Figure 8: Br evolutionary phase comparison for HARP 7959 at forecast hour 12. Rows corre￾spond to three evolutionary phases (emerging, steady, decaying); columns show ground truth (left), predicted field (center), and residual (right). Axes show Carrington longitude and latitude in degrees; images are displayed at the original HMI pixel resolution. A diverging red–blue col￾ormap centered at zero is used with a displ… view at source ↗
Figure 9
Figure 9. Figure 9: Same as Figure 8 but for HARP 8026. The model maintains structural fidelity [PITH_FULL_IMAGE:figures/full_fig_p014_9.png] view at source ↗
Figure 10
Figure 10. Figure 10: Br prediction for HARP 8206 (NOAA AR 13006) with pre-flare input only (input: 01:00–12:00 UT May 10). Rows: forecast hour 1 (13:00 UT, contains the X1.5 peak at 13:55 UT), hour 5 (17:00 UT, partial recovery ∼4 h after peak), and hour 12 (00:00 UT May 11, secondary degradation from continued AR activity). Columns show ground truth, predicted field, and residual. A ±500 G display range is used. 15 [PITH_FU… view at source ↗
Figure 11
Figure 11. Figure 11: Image RMSE [G] over the 12-hour forecast horizon (hours 1–12). Left: global RMSE. [PITH_FULL_IMAGE:figures/full_fig_p016_11.png] view at source ↗
Figure 12
Figure 12. Figure 12: SSIM(Br) and Pearson CC(Br) over the 12-hour forecast horizon (hours 1–12). Left (a): SSIM. Right (b): CC. 4.3 Magnetic-parameter prediction performance Beyond image-domain fidelity, we evaluate the model’s ability to preserve key magnetic￾parameter diagnostics derived from the predicted magnetic field components. These diagnostics—unsigned flux, magnetic pressure, shear angle, and field gradients—measure… view at source ↗
Figure 13
Figure 13. Figure 13: Per-hour relative error (%) for three magnetic-parameter diagnostics averaged over all [PITH_FULL_IMAGE:figures/full_fig_p018_13.png] view at source ↗
Figure 14
Figure 14. Figure 14: Absolute error of unsigned flux [G] over the 12-hour forecast horizon (hours 1–12). [PITH_FULL_IMAGE:figures/full_fig_p018_14.png] view at source ↗
Figure 15
Figure 15. Figure 15: Absolute error of magnetic pressure (Lorentz-force proxy) [G [PITH_FULL_IMAGE:figures/full_fig_p019_15.png] view at source ↗
Figure 16
Figure 16. Figure 16: Absolute error of shear angle [rad] over the 12-hour forecast horizon (hours 1–12). [PITH_FULL_IMAGE:figures/full_fig_p019_16.png] view at source ↗
Figure 17
Figure 17. Figure 17: Absolute error of |∇Br| [G/pixel] over the 12-hour forecast horizon (hours 1–12). Left: global evaluation. Right: masked-region evaluation [PITH_FULL_IMAGE:figures/full_fig_p020_17.png] view at source ↗
Figure 18
Figure 18. Figure 18: Absolute error of horizontal gradient |∇Bh| [G/pixel] over the 12-hour forecast horizon (hours 1–12). Left: global evaluation. Right: masked-region evaluation. 4.4.2 Statistical characteristics of orientation errors Pixel-level scatter analysis (Figs. 21 and 23) reveals strong linear correlation (CC>0.9) between predicted and reference orientation angles in strong-field regions. The tight clustering along… view at source ↗
Figure 19
Figure 19. Figure 19: Spatial distribution of orientation residuals for inclination [PITH_FULL_IMAGE:figures/full_fig_p021_19.png] view at source ↗
Figure 20
Figure 20. Figure 20: Same as Figure 19 but for HARP 8026 (NOAA AR 12953). [PITH_FULL_IMAGE:figures/full_fig_p022_20.png] view at source ↗
Figure 21
Figure 21. Figure 21: Orientation angle scatter comparison (θ and ϕ) in strong-field regions for HARP 7959 at forecast hours 6 and 12. Tight clustering along the diagonal demonstrates strong linear correlation. 23 [PITH_FULL_IMAGE:figures/full_fig_p023_21.png] view at source ↗
Figure 22
Figure 22. Figure 22: Gaussian fits of θ and ϕ residual distributions for HARP 7959 at forecast hours 6 and 12. Both approximate zero-mean normal distributions. 24 [PITH_FULL_IMAGE:figures/full_fig_p024_22.png] view at source ↗
Figure 23
Figure 23. Figure 23: Same as Figure 21 but for HARP 8026 [PITH_FULL_IMAGE:figures/full_fig_p025_23.png] view at source ↗
Figure 24
Figure 24. Figure 24: Same as Figure 22 but for HARP 8026 [PITH_FULL_IMAGE:figures/full_fig_p026_24.png] view at source ↗
Figure 25
Figure 25. Figure 25: Residual standard deviations (σ) for inclination θ and azimuth ϕ across 3,000 test sequences in strong-field regions. The relatively stable σ values across sequences indicate con￾sistent orientation prediction accuracy. 26 [PITH_FULL_IMAGE:figures/full_fig_p026_25.png] view at source ↗
Figure 26
Figure 26. Figure 26: Full temporal evolution of Br for HARP 7959 during the emerging phase. Top: 12-hour input sequence (GT). Middle: forecast hours 1–6 (GT vs. predicted). Bottom: forecast hours 7–12 (GT vs. predicted). 34 [PITH_FULL_IMAGE:figures/full_fig_p034_26.png] view at source ↗
Figure 27
Figure 27. Figure 27: Same as Figure 26 but for the [PITH_FULL_IMAGE:figures/full_fig_p035_27.png] view at source ↗
Figure 28
Figure 28. Figure 28: Same as Figure 26 but for the [PITH_FULL_IMAGE:figures/full_fig_p036_28.png] view at source ↗
Figure 29
Figure 29. Figure 29: Full temporal evolution of Br for HARP 8026 during the emerging phase. Layout is the same as [PITH_FULL_IMAGE:figures/full_fig_p037_29.png] view at source ↗
Figure 30
Figure 30. Figure 30: Same as Figure 29 but for the [PITH_FULL_IMAGE:figures/full_fig_p038_30.png] view at source ↗
Figure 31
Figure 31. Figure 31: Same as Figure 29 but for the [PITH_FULL_IMAGE:figures/full_fig_p039_31.png] view at source ↗
Figure 32
Figure 32. Figure 32: Bp evolutionary phase comparison for HARP 7959 at forecast hour 12. Layout is the same as [PITH_FULL_IMAGE:figures/full_fig_p041_32.png] view at source ↗
Figure 33
Figure 33. Figure 33: Bt evolutionary phase comparison for HARP 7959 at forecast hour 12. Layout is the same as [PITH_FULL_IMAGE:figures/full_fig_p041_33.png] view at source ↗
Figure 34
Figure 34. Figure 34: Bp evolutionary phase comparison for HARP 8026 at forecast hour 12. Layout is the same as [PITH_FULL_IMAGE:figures/full_fig_p042_34.png] view at source ↗
Figure 35
Figure 35. Figure 35: Bt evolutionary phase comparison for HARP 8026 at forecast hour 12. Layout is the same as [PITH_FULL_IMAGE:figures/full_fig_p042_35.png] view at source ↗
read the original abstract

Forecasting the dynamic evolution of solar magnetic fields is a critical technique for enabling space weather warnings. Addressing the limitations of existing methods in predicting all vector magnetic field components and in maintaining consistency with solar surface magnetic-field-related quantities, this study proposes a deep learning prediction method that integrates dynamic masks of active regions with multiple magnetic parameter constraints. By constructing a three-channel representation of vector magnetic fields, applying dynamic masks to enhance attention to strong-field regions, and incorporating multi-parameter magnetic parameter constraints, we developed an end-to-end short-term (12-hour) predictive model of solar vector magnetic field evolution. Using SDO/SHARP vector magnetogram data, the model predicts and analyses field evolution across all components. Quantitative evaluations demonstrate that our approach achieves horizon-averaged structural similarity index measure (SSIM) of 0.912 (per-hour range: 0.909--0.916) and correlation coefficient (CC) of 0.998 for the radial component Br (root-mean-square error (RMSE) 13.0--21.0 G); the horizontal components achieve Bphi SSIM 0.760--0.800 (CC 0.910--0.945, RMSE 38.5--50.0 G) and Btheta SSIM 0.728--0.750 (CC 0.895--0.920, RMSE 38.5--49.0 G). The model maintains unsigned magnetic flux prediction errors at 7.82% (95% confidence interval (CI): +/-0.11%). These results demonstrate strong image-domain performance together with consistency under the magnetic-parameter diagnostics used here, suggesting initial potential for supporting future space weather forecasting efforts.

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

3 major / 2 minor

Summary. The manuscript proposes an end-to-end deep-learning model for 12-hour-ahead prediction of the three-component vector magnetic field in solar active regions. The approach combines a three-channel input representation of SDO/SHARP vector magnetograms, dynamic masks that focus attention on strong-field pixels, and multi-parameter magnetic constraints (including unsigned flux) added to the training loss. On held-out data the model is reported to achieve horizon-averaged SSIM = 0.912 and CC = 0.998 (RMSE 13–21 G) for Br, lower but still usable SSIM/CC values for the horizontal components, and a mean unsigned-flux error of 7.82 % (95 % CI ±0.11 %).

Significance. If the performance numbers and physical-consistency claims are reproducible, the work would supply a concrete, data-driven baseline for short-term vector-field evolution that could be tested against existing physics-based or empirical forecasting pipelines. The explicit inclusion of magnetic-parameter constraints in the loss is a methodological strength that distinguishes the study from purely image-domain regression approaches.

major comments (3)
  1. [§3, §4] §3 (Methods) and §4 (Results): the explicit mathematical form of the multi-parameter loss terms, their relative weights, and the mechanism by which the constraints are enforced at inference time are not stated. Without these equations it is impossible to verify that the reported flux-error reduction does not arise from compensating errors among the three vector components.
  2. [§4.2] §4.2 (Quantitative evaluation): the 95 % CI on the 7.82 % flux error is given, but the manuscript does not specify whether the interval accounts for the number of independent active regions, temporal autocorrelation within each region, or multiple random seeds. This directly affects the load-bearing claim that the constraints produce statistically reliable consistency.
  3. [§3.1] §3.1 (Network architecture): no description is supplied of the backbone network, the precise definition of the dynamic masks, or the training/validation/test split ratios. These omissions prevent independent assessment of whether the quoted SSIM/CC values are architecture-dependent or genuinely attributable to the magnetic constraints.
minor comments (2)
  1. [§4.1] The per-hour SSIM ranges are reported only for Br; analogous ranges should be supplied for Bθ and Bφ to allow direct comparison of component-wise temporal stability.
  2. [Figure captions] Figure captions should explicitly state the number of active regions and the exact forecast horizons used to compute the quoted aggregate metrics.

Simulated Author's Rebuttal

3 responses · 0 unresolved

We thank the referee for the constructive and detailed comments, which highlight important aspects for improving reproducibility and statistical rigor. We address each major comment below and will revise the manuscript to incorporate clarifications and additional details where feasible.

read point-by-point responses

  1. Referee: [§3, §4] §3 (Methods) and §4 (Results): the explicit mathematical form of the multi-parameter loss terms, their relative weights, and the mechanism by which the constraints are enforced at inference time are not stated. Without these equations it is impossible to verify that the reported flux-error reduction does not arise from compensating errors among the three vector components.

    Authors: We agree that the explicit forms are necessary for verification. In the revised manuscript we will add the full mathematical definition of the composite loss (including the unsigned-flux term and any other magnetic-parameter penalties), the specific relative weights λ determined by cross-validation, and an explicit statement that the constraints operate exclusively during training. At inference the model performs unconstrained forward passes. We will also include a supplementary analysis of per-component residuals to demonstrate that the reported flux-error reduction is not produced by compensating errors across Br, Bθ and Bϕ. revision: yes

  2. Referee: [§4.2] §4.2 (Quantitative evaluation): the 95 % CI on the 7.82 % flux error is given, but the manuscript does not specify whether the interval accounts for the number of independent active regions, temporal autocorrelation within each region, or multiple random seeds. This directly affects the load-bearing claim that the constraints produce statistically reliable consistency.

    Authors: The reported 95 % CI was obtained via bootstrap resampling over the independent active regions in the held-out test set; five random seeds were used for training. We will add this description to §4.2. However, the original calculation did not explicitly block-bootstrap to account for temporal autocorrelation within each region. We will therefore revise the text to state the exact procedure used and to note this limitation; a full re-computation with clustered bootstrap would require additional experiments beyond the scope of a minor clarification and is left for future work. revision: partial

  3. Referee: [§3.1] §3.1 (Network architecture): no description is supplied of the backbone network, the precise definition of the dynamic masks, or the training/validation/test split ratios. These omissions prevent independent assessment of whether the quoted SSIM/CC values are architecture-dependent or genuinely attributable to the magnetic constraints.

    Authors: We will expand §3.1 with the requested details: the backbone is a U-Net architecture augmented with convolutional attention blocks; dynamic masks are binary maps generated by thresholding |Br| > 50 G (updated at each time step); and the data split is 70 % / 15 % / 15 % for training / validation / test, with active regions kept disjoint across splits to prevent leakage. These additions will allow readers to evaluate the contribution of the magnetic constraints independently of the architecture. revision: yes

Circularity Check

0 steps flagged

No circularity: standard DL training + independent test metrics

full rationale

The paper trains an end-to-end network on SDO/SHARP data with added loss terms for magnetic parameters and reports SSIM/CC/RMSE/flux error on held-out test magnetograms. No equation reduces a reported prediction to a fitted input by construction, no self-citation supplies a uniqueness theorem, and the evaluation metrics are external image and integral statistics computed after training. The derivation chain is therefore self-contained against external benchmarks.

Axiom & Free-Parameter Ledger

0 free parameters · 1 axioms · 0 invented entities

The central claim rests on standard machine-learning assumptions that a neural network can learn temporal evolution from image sequences when trained with a composite loss, and that the SDO/SHARP dataset provides representative examples of active-region evolution. No free parameters or invented entities are explicitly introduced in the abstract.

axioms (1)
  • domain assumption A neural network trained on sequences of vector magnetograms can learn to predict future states when the loss includes both image similarity and magnetic-parameter terms.
    Implicit in the end-to-end training description.

pith-pipeline@v0.9.1-grok · 5867 in / 1500 out tokens · 48626 ms · 2026-06-28T04:08:01.936563+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

25 extracted references · 15 canonical work pages

  1. [1]

    Aulanier, M

    G. Aulanier, M. Janvier, and B. Schmieder. The standard flare model in three dimensions. i. strong-to-weak shear transition in post-flare loops. Astronomy and Astrophysics, 543: A110, 2012. doi: 10.1051/0004-6361/201219311

    work page doi:10.1051/0004-6361/201219311 2012

  2. [2]

    Predicting the evolution of photospheric magnetic field in solar active regions using deep learning

    Liang Bai, Yi Bi, Bo Yang, Jun-Chao Hong, Zhe Xu, Zhen-Hong Shang, Hui Liu, Hai- Sheng Ji, and Kai-Fan Ji. Predicting the evolution of photospheric magnetic field in solar active regions using deep learning. Research in Astronomy and Astrophysics, 21(5):113,

  3. [3]

    doi: 10.1088/1674-4527/21/5/113

    work page doi:10.1088/1674-4527/21/5/113

  4. [4]

    Scheduled sampling for sequence prediction with recurrent neural networks

    Samy Bengio, Oriol Vinyals, Navdeep Jaitly, and Noam Shazeer. Scheduled sampling for sequence prediction with recurrent neural networks. In Advances in Neural Information Processing Systems, volume 28, pages 1171–1179, 2015

    2015

  5. [5]

    M. G. Bobra and S. Couvidat. Solar flare prediction using sdo/hmi vector magnetic field data with a machine-learning algorithm. Astrophysical Journal, 798:135, 2015

    2015

  6. [6]

    M. G. Bobra, X. Sun, J. T. Hoeksema, M. Turmon, Y. Liu, K. Hayashi, G. Barnes, and K. D. Leka. The Helioseismic and Magnetic Imager (HMI) Vector Magnetic Field Pipeline: SHARPs - Space-Weather HMI Active Region Patches.Sol. Phys., 289:3549–3578, Septem- ber 2014. doi: 10.1007/s11207-014-0529-3

    work page doi:10.1007/s11207-014-0529-3 2014

  7. [7]

    Sta- tistical analyses of solar active region in SDO/HMI magnetograms detected by unsuper- vised machine learning method DSARD

    Ruishuo Chen, Wutong Lu, Qi Hao, Yifan Meng, Pengfei Chen, and Chenxi Shi. Sta- tistical analyses of solar active region in SDO/HMI magnetograms detected by unsuper- vised machine learning method DSARD. arXiv preprint arXiv:2502.17990, feb 2025. doi: 10.48550/arXiv.2502.17990. URL https://arxiv.org/abs/2502.17990. 22 pages, 15 figures, submitted to ApJS

    work page doi:10.48550/arxiv.2502.17990 2025

  8. [8]

    Covas, N

    E. Covas, N. Peixinho, and J. Fernandes. Neural network forecast of the sunspot butterfly diagram. Solar Physics, 294(24):24, feb 2019. doi: 10.1007/s11207-019-1412-z. URL https://doi.org/10.1007/s11207-019-1412-z

    work page doi:10.1007/s11207-019-1412-z 2019

  9. [9]

    Transfer learning in spatial–temporal forecasting of the solar magnetic field

    Eurico Covas. Transfer learning in spatial–temporal forecasting of the solar magnetic field. Astronomische Nachrichten, 341(4), mar 2020. doi: 10.1002/asna.202013690. URL https://doi.org/10.1002/asna.202013690. First published: 30 March 2020

    work page doi:10.1002/asna.202013690 2020

  10. [10]

    Y. Cui, R. Li, L. Zhang, Y. He, and H. Wang. Correlation between solar flare pro- ductivity and photospheric magnetic field properties. i. maximum horizontal gradient, length of neutral line, number of singular points. Solar Physics, 237(1):45–59, 2006. doi: 10.1007/s11207-006-0077-6

    work page doi:10.1007/s11207-006-0077-6 2006

  11. [11]

    Y. Fan. MHD Equilibria and Triggers for Prominence Eruption, volume 415 of Astrophysics and Space Science Library, pages –. Springer, Cham, 2015. ISBN 978-3-319-10415-7. doi: 10.1007/978-3-319-10416-4\ 12. URL https://doi.org/10.1007/978-3-319-10416-4 12

    work page doi:10.1007/978-3-319-10416-4 2015

  12. [12]

    G. H. Fisher, D. J. Bercik, B. T. Welsch, and H. S. Hudson. Global forces in eruptive solar flares: The lorentz force acting on the solar interior. Solar Physics, 277:59, 2012

    2012

  13. [13]

    Prediction of the next solar rotation synoptic maps using an artificial intelligence–based surface flux transport model

    Hyun-Jin Jeong, Mingyu Jeon, Daeil Kim, Youngjae Kim, Ji-Hye Baek, Yong-Jae Moon, and Seonghwan Choi. Prediction of the next solar rotation synoptic maps using an artificial intelligence–based surface flux transport model. The Astrophysical Journal Supplement Series, 278(1):5, apr 2025. doi: 10.3847/1538-4365/adc447. 43

    work page doi:10.3847/1538-4365/adc447 2025

  14. [14]

    Prediction of time- evolving radial magnetic fields on the solar surface using deep learning

    Hyun-Jin Jeong, Stefaan Poedts, Haopeng Wang, Ekatarina Dineva, Panagiotis Gonidakis, Francesco Carella, George Miloshevich, Luis Linan, and Harim Lee. Prediction of time- evolving radial magnetic fields on the solar surface using deep learning. The Astrophysical Journal Letters, 989(2):L31, aug 2025. doi: 10.3847/2041-8213/adf5c7

    work page doi:10.3847/2041-8213/adf5c7 2025

  15. [15]

    Jiang, R

    J. Jiang, R. H. Cameron, and M. Sch¨ ussler. Effects of the scatter in sunspot group tilt angles on the large-scale magnetic field at the solar surface. Astrophysical Journal, 791(1): 5, 2014

    2014

  16. [16]

    Jiang, D

    J. Jiang, D. H. Hathaway, R. H. Cameron, S. K. Solanki, L. Gizon, and L. Upton. Magnetic flux transport at the solar surface. Space Science Reviews, 186(1-4):491–523, dec 2014. doi: 10.1007/s11214-014-0083-1. URL https://doi.org/10.1007/s11214-014-0083-1

    work page doi:10.1007/s11214-014-0083-1 2014

  17. [17]

    E. K. J. Kilpua, N. Lugaz, M. L. Mays, and M. Temmer. Forecasting the structure and orientation of earthbound coronal mass ejections. Space Weather, 17(4):498–526, 2019. doi: 10.1029/2018SW001944

    work page doi:10.1029/2018sw001944 2019

  18. [18]

    Adam: A method for stochastic optimization

    Diederik P Kingma and Jimmy Ba. Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980, 2014

    Pith/arXiv arXiv 2014

  19. [19]

    K. D. Leka and G. Barnes. Photospheric magnetic field properties of flaring versus flare- quiet active regions. ii. discriminant analysis. Astrophysical Journal, 595:1296, 2003

    2003

  20. [20]

    Georgoulis

    Francesco Pio Ramunno, Hyun-Jin Jeong, Stefan Hackstein, Andr´ e Csillaghy, Svyatoslav Voloshynovskiy, and Manolis K. Georgoulis. Magnetogram-to-magnetogram: Generative forecasting of solar evolution. arXiv preprint arXiv:2407.11659, jul 2024. doi: 10.48550/ arXiv.2407.11659. URL https://arxiv.org/abs/2407.11659. Conference paper accepted for oral present...

    arXiv 2024

  21. [21]

    Y. Su, A. Veronig, G. Holman, et al. Imaging coronal magnetic-field reconnection in a solar flare. Nature Physics, 9:489–493, 2013. doi: 10.1038/nphys2675

    work page doi:10.1038/nphys2675 2013

  22. [22]

    Flare-productive active regions

    Shin Toriumi and Haimin Wang. Flare-productive active regions. Living Reviews in Solar Physics, 16(1):3, 2019. doi: 10.1007/s41116-019-0019-7. URL https://link.springer.com/ article/10.1007/s41116-019-0019-7

    work page doi:10.1007/s41116-019-0019-7 2019

  23. [23]

    ADS Bibcode: 2013ApJ...773...93S The Astropy Collaboration, Price-Whelan, A

    L. Upton and D. H. Hathaway. Predicting the sun’s polar magnetic fields with a surface flux transport model. The Astrophysical Journal, 780(1):5, 2014. doi: 10.1088/0004-637X/ 780/1/5

    work page doi:10.1088/0004-637x/ 2014

  24. [24]

    H. Wu, Z. Yao, J. Wang, and M. Long. Motionrnn: A flexible model for video prediction with spacetime-varying motions. 2021

    2021

  25. [25]

    G. Yang, Y. Xu, W. D. Cao, et al. Photospheric shear flows along the magnetic neutral line of active region 10486 prior to an x10 flare. Astrophysical Journal, 617:L151, 2004. 44

    2004