Prediction of Magnetic Flux Evolution During Solar Active Region Emergence using Long Short-Term Memory Networks

Eren Dogan , Spiridon Kasapis , Sarang Patil , Jonas Tirona , John Stefan , Irina Kitiashvili , Mengjia Xu , Alexander Kosovichev

Authors on Pith no claims yet

Pith reviewed 2026-05-13 17:45 UTC · model grok-4.3

classification 🌌 astro-ph.SR astro-ph.IMphysics.space-ph

keywords activefluxmagfluxenc-decmagfluxlstmmagneticemergenceevolutionmodels

0 comments

The pith

Standard LSTM networks predict solar active region magnetic flux evolution 3-10 hours ahead from intensity and oscillation maps, outperforming encoder-decoder variants on held-out test regions.

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

Solar active regions are areas on the Sun where strong magnetic fields emerge, often triggering flares and storms that affect Earth. The study collected 1D time series data from 53 such regions plus quiet Sun areas, using maps of visible light intensity and vibration power at different frequencies. Two LSTM-based models were trained: a basic LSTM and a more complex encoder-decoder version. The simpler model proved more reliable, learning patterns that allowed it to forecast how magnetic flux would change over the next 12 hours, with useful accuracy starting 3 to 10 hours ahead on five test regions not seen during training.

Core claim

the simpler MagFluxLSTM, which can predict magnetic flux emergence 3-10 hours in advance within a 12-hour prediction window in both experimental and operational-type settings for the 5 testing active regions.

Load-bearing premise

That patterns learned from the 53 training active regions and their specific 30.66° field-of-view time series will generalize to new, unseen active regions without significant distribution shift or overfitting to the chosen frequency bands and preprocessing.

Figures

Figures reproduced from arXiv: 2604.03507 by Alexander Kosovichev, Eren Dogan, Irina Kitiashvili, John Stefan, Jonas Tirona, Mengjia Xu, Sarang Patil, Spiridon Kasapis.

**Figure 1.** Figure 1: The MagFluxEnc-Dec pipeline for magnetic flux prediction. The system extracts central tiles from SDO/HMI continuum intensity (Ic) and magnetic flux (Φm) maps, combines them with acoustic power maps to form feature tensor X ∈ RT ×N×(M+1) where T = 9 central tiles, N = 240 timesteps, and M = 4 power map frequencies. Sliding windows are generated by continuously sliding a window of length W = 110 timesteps ac… view at source ↗

**Figure 2.** Figure 2: Schematic overview of the MagFluxLSTM pipeline for magnetic flux prediction during active region emergence. MagFluxLSTM employs stacked LSTM structure instead of the encoder-decoder structure in [PITH_FULL_IMAGE:figures/full_fig_p007_2.png] view at source ↗

**Figure 3.** Figure 3: The Φm predictions for AR11726 obtained by the models presented in [PITH_FULL_IMAGE:figures/full_fig_p012_3.png] view at source ↗

**Figure 4.** Figure 4: Same as [PITH_FULL_IMAGE:figures/full_fig_p020_4.png] view at source ↗

**Figure 5.** Figure 5: Same as [PITH_FULL_IMAGE:figures/full_fig_p021_5.png] view at source ↗

**Figure 6.** Figure 6: Same as [PITH_FULL_IMAGE:figures/full_fig_p022_6.png] view at source ↗

**Figure 7.** Figure 7: Same as [PITH_FULL_IMAGE:figures/full_fig_p023_7.png] view at source ↗

read the original abstract

Solar active regions (ARs) are the primary drivers of space weather events, making their early prediction crucial for operational forecasting systems. We develop machine learning models capable of predicting the evolution of magnetic flux during AR emergence using 1D time series of the continuum intensity and solar oscillation power maps for 53 active regions and their surrounding quiet-Sun areas. Each observable is sampled over a fixed 30.66{\deg}x30.66{\deg} field of view. These observations capture the temporal evolution of each active region and serve as inputs for training and validation of our MagFluxLSTM and MagFluxEnc-Dec models. The MagFluxLSTM architecture implements a single-stage standard Long-Short Term Memory (LSTM) network. MagFluxEnc-Dec represents an LSTM encoder-decoder with teacher forcing. To test and evaluate the models' performance, we use the continuum intensity and oscillation power maps (calculated for several frequency bands from Doppler velocity) as input to predict the magnetic flux. Among the top 100 hyperparameter configurations ranked by validation derivative RMSE, 98% correspond to MagFluxLSTM, compared to only 2% for MagFluxEnc-Dec. Thus, although the MagFluxEnc-Dec architecture has higher model complexity, it leads to poorer generalization to ARs outside the training set and less stable training than the simpler MagFluxLSTM, which can predict magnetic flux emergence 3-10 hours in advance within a 12-hour prediction window in both experimental and operational-type settings for the 5 testing active regions.

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.

Desk Editor's Note private letter to a colleague

The simpler LSTM beats the encoder-decoder on flux prediction from intensity and oscillation data, but five test regions give weak evidence for reliable generalization.

read the letter

The one thing to know is that the simpler MagFluxLSTM outperforms the encoder-decoder version, with 98 percent of top hyperparameter runs favoring it, yet the whole evaluation rests on only five held-out active regions. The paper trains these models on 1D time series of continuum intensity plus multi-band oscillation power sampled inside a fixed 30.66-degree field of view for 53 training regions. It reports usable 3-to-10-hour-ahead predictions inside a 12-hour window on the test cases. That specific input combination and the clear win for the basic architecture are the main increments over prior solar ML work. The result is practically useful because it shows the more complex model generalizes worse and trains less stably, which is worth knowing for anyone building similar forecasters. The soft spots are the small test sample and missing details. Five regions supply limited statistical power, with no reported variance across splits, no diversity metrics on the test ARs, and no explicit checks for distribution shift or temporal leakage. The abstract also omits error bars and baseline comparisons. These gaps make it hard to judge how well the patterns will hold for new regions. Readers working on applied machine learning for space weather or solar active-region studies would get value from the architecture comparison and input choices. The work shows honest engagement with the modeling trade-offs and deserves peer review so the data handling and generalization claims can be examined more closely. I would send it out rather than desk reject.

Circularity Check

0 steps flagged

No significant circularity in the LSTM time-series forecasting pipeline

full rationale

The paper trains standard LSTM architectures (MagFluxLSTM and MagFluxEnc-Dec) on 1D time series of continuum intensity and oscillation power maps drawn from 53 active regions, then evaluates forecasts on a held-out set of 5 regions. No algebraic derivation, parameter fitting, or self-referential definition reduces the claimed 3-10 hour predictions to the training inputs by construction. Hyperparameter ranking on validation RMSE introduces ordinary selection bias but does not create a tautology; the network weights are learned from data rather than presupposing the target flux values. No load-bearing self-citations or uniqueness theorems appear in the provided text. The workflow is therefore a conventional supervised learning setup whose outputs are not forced by the inputs.

Axiom & Free-Parameter Ledger

2 free parameters · 2 axioms · 0 invented entities

The central claim rests on standard neural network training assumptions plus domain-specific choices about which solar observables capture emergence physics; no new physical entities are postulated.

free parameters (2)

LSTM hyperparameters (hidden size, layers, learning rate, etc.)
Top 100 configurations selected by validation derivative RMSE; these are fitted choices that directly affect reported performance.
Frequency bands for oscillation power maps
Chosen inputs whose exact band definitions and number are not detailed in the abstract but affect the input features.

axioms (2)

domain assumption LSTM networks can learn temporal dependencies in solar time series that correlate with future magnetic flux changes
Invoked by the choice of architecture and the claim that the model predicts emergence.
domain assumption The 30.66° field-of-view sampling and 1D time series representation preserve sufficient information for flux prediction
Stated in the data description section of the abstract.

pith-pipeline@v0.9.0 · 5606 in / 1551 out tokens · 65696 ms · 2026-05-13T17:45:35.665163+00:00 · methodology

discussion (0)

Reference graph

Works this paper leans on

28 extracted references · 28 canonical work pages

[1]

Barnes, G., Birch, A., Leka, K., Braun, D.: 2014, Helioseismology of pre-emerging active regions. III. Statistical analysis.The Astrophysical Journal786,

work page 2014
[2]

Bergstra, J., Yamins, D., Cox, D.D., et al.: 2013, Hyperopt: A Python Library for Optimizing the Hyperparameters of Machine Learning Algorithms.SciPy13,

work page 2013
[3]

Birch, A., Braun, D., Leka, K., Barnes, G., Javornik, B.: 2012, Helioseismology of pre-emerging active regions. II. Average emergence properties.The Astrophysical Journal762,

work page 2012
[4]

Bobra, M.G., Ilonidis, S.: 2016, Predicting coronal mass ejections using machine learning methods.The Astrophysical Journal821,

work page 2016
[5]

Camporeale, E.: 2020, ML-Helio: An Emerging Community at the Intersection Between Heliophysics and Machine Learning.Journal of Geophysical Research (Space Physics) 125, e27502. DOI. ADS. Georgoulis, M.K., Bloomfield, D.S., Piana, M., Massone, A.M., Soldati, M., Gallagher, P.T., Pariat, E., Vilmer, N., Buchlin, E., Baudin, F., et al.: 2021, The flare likel...

work page 2020
[6]

Grigor’ev, V., Ermakova, L., Khlystova, A.: 2011, The dynamics of photospheric line-of-sight velocities in emerging active regions.Astronomy reports55,

work page 2011
[7]

Hartlep, T., Kosovichev, A.G., Zhao, J., Mansour, N.N.: 2011, Signatures of emerging subsurface structures in acoustic power maps of the Sun.Solar Physics268,

work page 2011
[8]

Astrophysics and Space Science370,

Hegde, M.: 2025, Predicting CME speed at 20 R ⊙using machine learning approaches. Astrophysics and Space Science370,

work page 2025
[9]

Hochreiter, S., Schmidhuber, J.: 1997, Long short-term memory.Neural computation9,

work page 1997
[10]

Hoeksema, J.T., Liu, Y., Hayashi, K., Sun, X., Schou, J., Couvidat, S., Norton, A., Bobra, M., Centeno, R., Leka, K.D., Barnes, G., Turmon, M.: 2014, The Helioseismic and Magnetic Imager (HMI) Vector Magnetic Field Pipeline: Overview and Performance.Solar Physics 289,

work page 2014
[11]

Hua, Y., Zhao, Z., Li, R., Chen, X., Liu, Z., Zhang, H.: 2019, Deep learning with long short-term memory for time series prediction.IEEE Communications Magazine57,

DOI.URL. Hua, Y., Zhao, Z., Li, R., Chen, X., Liu, Z., Zhang, H.: 2019, Deep learning with long short-term memory for time series prediction.IEEE Communications Magazine57,

work page 2019
[12]

Manuscript under review

Hutchins, T., Kasapis, S., Farooki, H., Cuesta, M.E., Khoo, L., Pak, S., Czarnota, R., Rankin, J., Szalay, J., Xu, Z., Livadiotis, G., Sarlis, N., McComas, D.: 2026, Solar Energetic Particle Prediction in the Inner Heliosphere Using Deep Learning and PSP/IS ⊙IS Data.JGR: Machine Learning and Computation. Manuscript under review. Ilonidis, S., Zhao, J., Ha...

work page 2026
[13]

Ilonidis, S., Zhao, J., Kosovichev, A.: 2011, Detection of emerging sunspot regions in the solar interior.Science333,

work page 2011
[14]

Jiao, Z., Sun, H., Wang, X., Manchester, W., Gombosi, T., Hero, A., Chen, Y.: 2020, Solar flare intensity prediction with machine learning models.Space weather18, e2020SW002440. Kasapis, S., Kitiashvili, I.N., Kosovichev, A.G., Stefan, J.T., Apte, B.: 2023, Predicting the Emergence of Solar Active Regions Using Machine Learning.Proceedings of the Internat...

work page 2020
[15]

Kasapis, S., Kitiashvili, I.N., Kosovichev, A.G., Stefan, J.T.: 2025, Prediction of Intensity Variations Associated with Emerging Active Regions using Helioseismic Power Maps and Machine Learning.The Astrophysical Journal Supplement Series280,

work page 2025
[16]

Kasapis, S., Hosseinzadeh, P., Whitman, K., Egeland, R., Georgoulis, M., Vourlidas, A., Papaioannou, A., Lavasa, E., Anastasiadis, A., Giannopoulos, G., Mu˜ noz-Jaramillo, A., Poduval, B., Kitiashvili, I.N., Kosovichev, A.G., Sadykov, V., Filali Boubrahimi, S., Hutchins, T.T., Farooki, H.A., Cuesta, M.E., Khoo, L.Y., Pak, S., Czarnota, R., Rankin, J.S., S...

work page arXiv 2026
[17]

DOI. ADS. Kosovichev, A.G., Duvall, T.L., Scherrer, P.H.: 1999, Time-distance helioseismology.Advances in Space Research24,

work page 1999
[18]

DOI. ADS. Leka, K.D., Barnes, G., Birch, A.C., Gonzalez-Hernandez, I., Dunn, T., Javornik, B., Braun, D.C.: 2013, Helioseismology of Pre-emerging Active Regions. I. Overview, Data, and Target Selection Criteria.ApJ762,

work page 2013
[19]

DOI. ADS. Li, L., Jamieson, K., Rostamizadeh, A., Gonina, E., Ben-Tzur, J., Hardt, M., Recht, B., Talwalkar, A.: 2020, A system for massively parallel hyperparameter tuning.Proceedings of machine learning and systems2,

work page 2020
[20]

Liaw, R., Liang, E., Nishihara, R., Moritz, P., Gonzalez, J.E., Stoica, I.: 1807, Tune: a research platform for distributed model selection and training (2018).ArXiv:1807.05118. Liaw, R., Liang, E., Nishihara, R., Moritz, P., Gonzalez, J.E., Stoica, I.: 2018, Tune: A Research Platform for Distributed Model Selection and Training.arXiv preprint arXiv:1807....

work page Pith review arXiv 2018
[21]

Pesnell, W.D., Thompson, B.J., Chamberlin, P.: 2012,The solar dynamics observatory (SDO), Springer. Roy, S., Schmude, J., Lal, R., Gaur, V., Freitag, M., Kuehnert, J., van Kessel, T., Hegde, D.V., Mu˜ noz-Jaramillo, A., Jakubik, J., et al.: 2025a, Surya: Foundation model for heliophysics. arXiv preprint arXiv:2508.14112. Roy, S., Hegde, D.V., Schmude, J.,...

work page arXiv 2012
[22]

Schrijver, C.J.: 2009, Driving major solar flares and eruptions: A review.Advances in Space Research43,

DOI.URL. Schrijver, C.J.: 2009, Driving major solar flares and eruptions: A review.Advances in Space Research43,

work page 2009
[23]

Monthly Notices of the Royal Astronomical Society533,

Schunker, H., Roland-Batty, W., Birch, A.C., Braun, D.C., Cameron, R.H., Gizon, L.: 2024, A flux-independent increase in outflows prior to the emergence of active regions on the Sun. Monthly Notices of the Royal Astronomical Society533,

work page 2024
[24]

Singh, N.K., Raichur, H., Brandenburg, A.: 2016, High-wavenumber solar f-mode strengthening prior to active region formation.The Astrophysical Journal832,

Sherstinsky, A.: 2020, Fundamentals of recurrent neural network (RNN) and long short-term memory (LSTM) network.Physica D: Nonlinear Phenomena404, 132306. Singh, N.K., Raichur, H., Brandenburg, A.: 2016, High-wavenumber solar f-mode strengthening prior to active region formation.The Astrophysical Journal832,

work page 2020
[25]

Stefan, J.T., Kosovichev, A.G.: 2023, Exploring the Connection between Helioseismic Travel Time Anomalies and the Emergence of Large Active Regions during Solar Cycle 24.The Astrophysical Journal948,

work page 2023
[26]

Waidele, M., Roth, M., Singh, N., K¨ apyl¨ a, P.: 2023, On Strengthening of the Solar f-Mode Prior to Active Region Emergence Using the Fourier-Hankel Analysis.Solar Physics298,

Tirona, J., Patil, S., Kasapis, S., Dogan, E., Stefan, J., Kitiashvili, I.N., Kosovichev, A.G., Xu, M.: 2026, Forecasting Continuum Intensity for Solar Active Region Emergence Prediction using Transformers.arXiv preprint arXiv:2601.13144. Waidele, M., Roth, M., Singh, N., K¨ apyl¨ a, P.: 2023, On Strengthening of the Solar f-Mode Prior to Active Region Em...

work page arXiv 2026
[27]

17 Appendix A

SOLA: solarphysicsjournal.tex; 7 April 2026; 0:16; p. 17 Appendix A. Hyperparameter Optimization Hyperparameter optimization was performed using Ray Tune using the Hyperopt search algorithm and the Asynchronous Successive Halving Algorithm (ASHA) for early-stopping of underperforming trials. The optimization objective was set to minimize the validation ro...

work page 2026
[28]

l e a r n i n g _ r a t e

Listing 1: Ray Tune + Hyperopt search configuration for LSTM hyperparameter optimization. s e a r c h _ s p a c e = { " l e a r n i n g _ r a t e " : hp . l o g u n i f o r m ( " l e a r n i n g _ r a t e " , log (1 e -5) , log (1 e -2) ) , " h i d d e n _ s i z e " : hp . choice ( " h i d d e n _ s i z e " , [2 , 4 , 8 , 16 , 32 , 64 , 128]) , " n u m _ ...

work page 2026