Pith. sign in

REVIEW 1 major objections 5 minor 1 cited by

A multi-task deep-learning model that combines magnetic-field, X-ray, and historical proton-flux data improves 24-hour-ahead forecasts of solar energetic particle events.

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

2026-08-02 11:27 UTC pith:GDB7R5QO

load-bearing objection Solid dataset and benchmark extension with honest limitations, but the headline forecast skill numbers rest on a non-operational i.i.d. split and need walk-forward confirmation. the 1 major comments →

arxiv 2606.14440 v2 pith:GDB7R5QO submitted 2026-06-12 astro-ph.SR physics.space-ph

Realtime forecasting of solar energetic particle event and proton flux using multi-source solar observations and multi-task deep learning

Yian Yu , Yang Chen , Lulu Zhao , Kathryn Whitman , Ward Manchester , Tamas Gombosi This is my paper
classification astro-ph.SR physics.space-ph
keywords solar energetic particlesSEP forecastingmulti-task deep learningspace weatherproton fluxsoft X-ray fluxactive region magnetogramsrare-event prediction
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.

This paper attempts to show that 24-hour-ahead forecasts of solar energetic particle (SEP) events—sudden increases in >10 MeV proton flux that endanger spacecraft and astronauts—get better when a machine-learning model is fed several complementary kinds of solar observations at once rather than any single data source. The authors construct a dataset of 14,464 non-overlapping day-long windows covering 1986–2025, labeled positive when the next 24 hours contain an event exceeding 10 pfu, and train SEPNET-Ov2, a multi-task model that simultaneously predicts event occurrence, future proton-flux level, and future soft X-ray flux. Their best configuration, combining magnetic-field parameters, historical proton flux, and X-ray flux, reaches a test AUC of 0.925 and a true-skill statistic of 0.666, clearly above the proton-flux-only model (0.845) and above the earlier SEPNET-O on the same benchmark. A fair reader would take this as evidence that broadening historical magnetic coverage and adding radiative and flux precursors is a productive route for operational space-weather forecasting.

Core claim

The paper's central claim is that SEPNET-Ov2—a multi-task deep-learning architecture that takes two consecutive 24-hour summaries of heterogeneous solar observations and jointly predicts SEP occurrence, future >10 MeV proton flux, and future soft X-ray flux—delivers operationally stronger 24-hour-ahead SEP forecasts than any single predictor group, with the best combination (SMHARP magnetic parameters + historical proton flux + X-ray flux) yielding median test AUC 0.9253, TSS 0.6660, HSS 0.6642, F1 0.6806 and FAR 0.3206. It further claims that the improved performance comes from both the dataset and the model: extending magnetic parameters back to 1996 via SMARP alignment and adding flux-bas

What carries the argument

The central object is SEPNET-Ov2's hybrid multi-task architecture. Each sample is a sequence of length L=2 containing min/mean/max summaries of all predictors over the current and previous 24-hour windows; a bidirectional LSTM followed by a Transformer encoder processes the temporal sequence, while a multilayer perceptron processes the most recent vector, and the fused representation feeds three heads: binary SEP classification, proton-flux regression, and soft X-ray regression. The supporting mechanism is the SMHARP dataset, created by aligning SHARP and SMARP active-region magnetic products through feature-wise multivariate regression on overlapping observations, which extends magnetic pre

Load-bearing premise

The load-bearing premise is that randomly splitting the 14,464 daily samples into train/test sets mimics realtime forecasting; since adjacent windows share overlapping event histories, the reported scores are liable to be optimistic until a temporally ordered evaluation confirms them.

What would settle it

Run SEPNET-Ov2 under a strict chronological split—train on all windows before a cutoff date, predict each subsequent day using only past data—and compare with a persistence baseline that warns whenever the previous 24-hour mean >10 MeV proton flux exceeds some threshold. If the multi-source model's AUC/HSS advantage over persistence nearly disappears, the central claim that magnetic and X-ray inputs add skill is not supported.

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

If this is right

  • If the central claim holds, operational SEP warning systems can gain skill at a fixed 24-hour horizon by combining magnetic, proton-flux, and X-ray summaries, without relying on CME parameters.
  • The reported gains imply that historical proton-flux persistence is a strong baseline: any future model should be compared against it, and the margin over it is the meaningful skill contribution.
  • Extending magnetic predictors back to 1996 means the model can be trained and evaluated across more solar cycles, which should make claimed skill more stable across activity levels.
  • Because the best configuration keeps the false alarm ratio near one third while catching roughly 68% of events, the model is closer to being usable in a real warning service, though false alarms remain a stated limitation.
  • The multi-task result suggests the auxiliary flux-regression heads act as a regularizer for the rare-event classification, a design that can transfer to other sparse space-weather predictions.

Where Pith is reading between the lines

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

  • If a strict chronological (walk-forward) evaluation were applied, the reported gap over the proton-flux-only baseline might shrink because consecutive windows share event histories; the paper itself flags this, so the operational skill should be read as an upper bound until such an evaluation is done.
  • The SHARP/SMARP alignment strategy could be stress-tested by checking whether a model trained only on the post-2010 overlap period retains skill when tested pre-2010; the paper reports alignment correlations but not forecast-level transfer validation.
  • Since CME variables added little, operational pipelines may drop them to reduce data latency; a testable extension is to measure forecast skill in near-real-time conditions where CME catalogs are incomplete for hours after an eruption.
  • If auxiliary regression targets improve classification, an even stronger design might use higher-cadence (hourly) flux inputs or add proton anisotropy and radio burst streams, which the summarized daily windows currently discard.

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

1 major / 5 minor

Summary. The paper presents SEPNET-PRISM/SEPNET-Ov2, a multi-task deep-learning framework for 24-hour-ahead forecasting of operational solar energetic particle (SEP) events (defined by >10 MeV proton flux exceeding 10 pfu). The framework integrates heterogeneous predictors: SMHARP magnetic parameters (SHARP+SMARP aligned), flare properties, CME catalog data (CDAWDONKI), GOES XRSB soft X-ray flux, and historical >10 MeV proton flux. On the CLEAR benchmark dataset (14,464 non-overlapping 24-hour windows, 650 positives), the authors report that the best configuration (S+PF+X) achieves median test AUC 0.925, TSS 0.666, HSS 0.664, outperforming the proton-flux-only model (AUC 0.845) and the earlier SEPNET-O model on the newly aligned dataset. The paper also evaluates auxiliary regression targets (future maximum log proton flux and XRSB flux) and a permutation-importance analysis. The authors explicitly acknowledge in the Discussion that the experiments use random i.i.d. splits and that temporally ordered testing should be emphasized in future operational evaluation.

Significance. If the reported evaluation was operationally valid, the contribution would be significant: it assembles a unified multi-cycle dataset with extended magnetic-field coverage back to 1996, demonstrates that multi-source integration improves over single-source models, releases code and data, and targets a practically relevant 24-hour warning horizon. The multi-task architecture and the SMHARP alignment procedure are useful methodological building blocks. However, the central empirical claim is not yet established at the claimed strength because the evaluation protocol — random i.i.d. splitting of temporally autocorrelated daily windows with overlapping event labels — is acknowledged by the authors to be mismatched to realtime forecasting, and no persistence or climatology baseline is reported. The proton-flux-only model's high AUC is consistent with persistence-like behavior, so the added value of multi-source integration over a trivial ongoing-event detector remains unquantified. The paper's strengths are the dataset construction, the reproducibility commitment, and the transparency about limitations; its weakness is that the headline performance numbers are likely upper bounds on rea

major comments (1)
  1. [Discussion and conclusions] The manuscript states the framework 'achieves strong skill while maintaining low false-positive rates' and later calls the results 'an important step toward operational deployment.' Given the acknowledged i.i.d. split and the missing temporal baseline, these conclusions overstate what has been demonstrated. The paper should explicitly qualify the headline numbers as upper bounds until a temporally ordered evaluation is performed, and should temper the operational-deployment language accordingly.
minor comments (5)
  1. [Abstract/Introduction] The abstract uses 'SEPNET-PRISM' while the body uses 'SEPNET-Ov2'; please harmonize the naming.
  2. [Results] Typographical errors: 'an finally' in the first paragraph of Results, 'catelog' in Methods, and 'Meta-probabilities ... are also be incorporated' in the architecture description.
  3. [Table 2] The R² values for some configurations are strongly negative (e.g., PF-only log10 proton flux R² = -0.5048). This is worth a sentence of interpretation, as it indicates the auxiliary regression task performs worse than the mean predictor for those feature sets, which bears on the claim that auxiliary tasks help the shared representation.
  4. [References] The paper cites the earlier SEPNET paper as '2512.12786' and several 2025–2026 data releases. Please verify all citations are in a consistent and accessible format, and consider adding a reference to recent SEP forecasting benchmarks/evaluation protocols for context.
  5. [Figure 1] The representative test-set predictions in Figure 1 are from one seed and one configuration; it would be clearer to show a time-ordered sequence with event intervals marked, so the reader can see whether the model is detecting onsets or ongoing events.

Circularity Check

0 steps flagged

No circular derivation found; the evaluation-split limitation is a validation-realism concern, not a tautology.

full rationale

The paper's derivation chain is empirical and self-contained: it assembles a dataset from external catalogs (CLEAR, GOES, SHARP/SMARP, DONKI/CDAW), defines the positive label from CLEAR SEP events (future >10 MeV proton flux >10 pfu), and evaluates trained models on held-out samples. The use of historical >10 MeV proton flux as an input while the target is future >10 MeV flux is legitimate persistence-style forecasting, not a definitional identity: the input summarizes the preceding 24-hour window and the label covers the subsequent 24-hour window, so the prediction does not reduce to the input by construction. The auxiliary regression of future maximum log proton flux similarly targets a different time interval than the input summary. The only self-citations (SEPNET ref. 22, and CLEAR/FetchSEP refs. 24/31 from co-authors) provide architectural continuity and benchmark data, but the central improvement claim is tested anew in this paper on its own test set, so no load-bearing result is imported purely by citation. The manuscript explicitly acknowledges that 'the present experiments use random i.i.d. splits, whereas future operational evaluation should emphasize temporally ordered testing'; that is an evaluation-protocol limitation that could affect the reported skill scores, but it is not a circular derivation of the claimed prediction from its inputs.

Axiom & Free-Parameter Ledger

4 free parameters · 4 axioms · 0 invented entities

This is an empirical machine-learning paper, so its 'free parameters' are fitted alignment mappings, a tuned decision threshold, and trained network weights rather than physical constants. The load-bearing assumptions are the ones the paper itself flags (i.i.d. evaluation) plus the generalization of regression alignments outside their fitted overlap windows. No new physical entities are introduced: SMHARP and CDAWDONKI are merged observational archives, not new forces, particles, or conserved quantities.

free parameters (4)
  • SMARP->SHARP reconstruction regression coefficients = not reported (per-parameter multivariate regression; overlap-period correlations 0.84-0.98)
    The unified SMHARP magnetic features used as S/S_AR inputs are produced by this fitted mapping; reconstruction error bounds the information content of all pre-2010 magnetic predictors.
  • CDAW->DONKI linear mapping coefficients = not reported
    All CME-catalog features (C) are generated by applying this mapping to CDAW events back to 1996; mapping errors propagate into the 1996-2010 portion of the C feature set.
  • Decision threshold (probability cutoff) = not reported (selected on validation to maximize min(TSS,HSS))
    All classification metrics in Tables 1 and the SEPNET-O comparison are evaluated at a validation-tuned threshold rather than the default 0.5; the reported FAR/HSS are conditional on this choice.
  • Network weights and Optuna hyperparameters = not reported
    The central predictions are outputs of trained weights, and per-configuration Optuna tuning (regularization, focal loss, loss weights) adds model-selection freedom beyond the five-seed median.
axioms (4)
  • domain assumption Temporal exchangeability of the 14,464 24-hour windows under random splits
    The i.i.d. evaluation assumes adjacent windows carry independent information; the authors themselves flag this as unrealistic for operational use in the Discussion.
  • domain assumption SMARP->SHARP alignment generalizes beyond the May-October 2010 overlap period
    The regression mapping is fitted on a six-month overlap and then applied to 1996-2010 data; no out-of-overlap validation is shown in the main text.
  • domain assumption 24-hour min/mean/max summaries preserve the predictive signal in the raw time series
    The entire feature representation discards sub-daily timing and event structure; justified only by the authors' prior SEPNET study (ref 22).
  • domain assumption CLEAR/GOES >10 MeV event list is reliable ground truth
    Labels, the PF input, and the evaluation benchmark all derive from the same GOES >10 MeV proton channel via FetchSEP/CLEAR; event start/end boundaries are taken as authoritative.

pith-pipeline@v1.3.0-alltime-deepseek · 10478 in / 16759 out tokens · 174486 ms · 2026-08-02T11:27:22.006014+00:00 · methodology

0 comments
read the original abstract

Solar energetic particle (SEP) events, defined by proton flux exceeding 10 pfu in the > 10 MeV channel, pose major risks to spacecraft operations, astronaut safety, and high-latitude aviation. Due to the complexity and rarity of SEP events, reliable operational SEP forecasting remains an important challenge in space weather. Here we present a novel 24-hour-ahead realtime forecasting framework, SEPNET-PRISM, based on a multi-task learning structure and a thoroughly constructed list of features from multiple sources spanning multiple solar cycles, that jointly predicts SEP event occurrence and future proton and soft X-ray fluxes. SEPNET-PRISM extends the earlier-introduced SEPNET-based models by integrating a broader range of solar observations, including active-region magnetic parameters from SHARP and SMARP, solar-flare information, coronal mass ejections, soft X-ray flux, and historical > 10 MeV proton flux. As compared with SEPNET, the inclusion of SMARP data expands the temporal coverage of magnetic-field predictors to earlier solar cycles, while flux-based inputs provide additional precursor information. Evaluation on the CLEAR SEP benchmark dataset shows improved classification performance over the earlier SEPNET-O (operational version of SEPNET) on the newly aligned dataset. The best operational model is obtained when magnetic, radiative, and proton-flux predictors are combined, highlighting the value of expanded historical coverage and complementary precursor information for improving realtime SEP forecasting.

Figures

Figures reproduced from arXiv: 2606.14440 by Kathryn Whitman, Lulu Zhao, Tamas Gombosi, Ward Manchester, Yang Chen, Yian Yu.

Figure 1
Figure 1. Figure 1: Representative test-set predictions of SEPNET-Ov2 for the S+PF+F feature configuration under one random seed. 3/12 [PITH_FULL_IMAGE:figures/full_fig_p003_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Comparison of classification performance among the earlier SEPNET-O results reported previously, SEPNET-O evaluated on the newly aligned dataset, and SEPNET-Ov2 under the S+F feature configuration. 5/12 [PITH_FULL_IMAGE:figures/full_fig_p005_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Relative feature importance of predictor groups for operational SEP forecasting derived from permutation importance in repeated, class-balanced random forest models. The pie charts provide a compact summary of the aggregate importance of predictors by data source and variable family. For each predictor, importance was aggregated across the minimum, average, and maximum values computed over the preceding no… view at source ↗
Figure 4
Figure 4. Figure 4: Schematic of the preprocessing workflow used to convert heterogeneous solar observations into supervised-learning samples. Data from each source were aggregated into fixed, non-overlapping 24-hour historical windows, and each predictor variable was represented by its minimum, mean, and maximum values within the window. Prediction targets were defined over the subsequent 24-hour window and included a binary… view at source ↗
Figure 5
Figure 5. Figure 5: Architecture of SEPNET-Ov2. Each sample is represented as a sequence of two summarized 24-hour predictor vectors. A temporal branch processes the sequence using a bidirectional LSTM followed by a Transformer encoder, while a parallel tabular branch encodes the most recent predictor vector. The resulting representations are fused and passed to task-specific heads for operational SEP classification and auxil… view at source ↗

discussion (0)

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

Forward citations

Cited by 1 Pith paper

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score.

  1. SEP-PRISM Data: A multi-source dataset for solar energetic particle forecasting

    astro-ph.SR 2026-07 conditional novelty 6.0

    A curated multi-source dataset with 14,464 daily samples (650 operational SEP positives) for 24-hour-ahead solar energetic particle forecasting, extending magnetic and CME predictors back to 1996 via regression alignment.

Reference graph

Works this paper leans on

34 extracted references · 7 canonical work pages · cited by 1 Pith paper

  1. [1]

    P.et al.The economic impact of space weather: Where do we stand?Risk Analysis37, 206–218, DOI: 10.1111/risa.12765 (2017)

    Eastwood, J. P.et al.The economic impact of space weather: Where do we stand?Risk Analysis37, 206–218, DOI: 10.1111/risa.12765 (2017)

    doi:10.1111/risa.12765 2017

  2. [2]

    InTriennial Earth-Sun Summit(2024)

    Whitman, K.et al.A multi-year effort to forward the validation of solar energetic particle models. InTriennial Earth-Sun Summit(2024)

    2024

  3. [3]

    Solar energetic particle variations.Adv

    Reames, D. Solar energetic particle variations.Adv. Space Res.34, 381–390, DOI: 10.1016/j.asr.2003.02.046 (2004). Solar Variability and Climate Change

    doi:10.1016/j.asr.2003.02.046 2003

  4. [4]

    Y ., De Angelis, G

    Kim, M.-H. Y ., De Angelis, G. & Cucinotta, F. A. Probabilistic assessment of radiation risk for astronauts in space missions. Acta Astronaut.68, 747–759, DOI: 10.1016/j.actaastro.2010.08.035 (2011)

    doi:10.1016/j.actaastro.2010.08.035 2010

  5. [5]

    & Giacalone, J

    Desai, M. & Giacalone, J. Large gradual solar energetic particle events.Living Rev. Sol. Phys.13, 3, DOI: 10.1007/ s41116-016-0002-5 (2016)

    2016

  6. [6]

    & Dalla, S

    Klein, K.-L. & Dalla, S. Acceleration and propagation of solar energetic particles.Space Sci. Rev.212, 1107–1136, DOI: 10.1007/s11214-017-0382-4 (2017)

    doi:10.1007/s11214-017-0382-4 2017

  7. [7]

    & Shea, M

    Smart, D. & Shea, M. PPS76: A computerized event mode solar proton forecasting technique. InNOAA Solar-Terrestrial Predictions Proceedings, vol. 1, 406–427 (1979)

    1979

  8. [8]

    Laurenza, M.et al.A technique for short-term warning of solar energetic particle events based on flare location, flare size, and evidence of particle escape.Space Weather.7, DOI: 10.1029/2007SW000379 (2009)

    doi:10.1029/2007sw000379 2009

  9. [9]

    V .et al.A new field line advection model for solar particle acceleration.The Astrophys

    Sokolov, I. V .et al.A new field line advection model for solar particle acceleration.The Astrophys. J.616, L171, DOI: 10.1086/426812 (2004)

    doi:10.1086/426812 2004

  10. [10]

    & Riley, P

    Luhmann, J., Ledvina, S., Krauss-Varban, D., Odstrcil, D. & Riley, P. A heliospheric simulation-based approach to SEP source and transport modeling.Adv. Space Res.40, 295–303, DOI: 10.1016/j.asr.2007.03.089 (2007)

    work page doi:10.1016/j.asr.2007.03.089 2007

  11. [11]

    Hu, J., Li, G., Ao, X., Zank, G. P. & Verkhoglyadova, O. Modeling particle acceleration and transport at a 2-D CME-driven shock.J. Geophys. Res. Space Phys.122, 10,938–10,963, DOI: 10.1002/2017JA024077 (2017)

    work page doi:10.1002/2017ja024077 2017

  12. [12]

    Zhao, L.et al.Solar wind with field lines and energetic particles (SOFIE) model: Application to historical solar energetic particle events.Space Weather.22, e2023SW003729, DOI: 10.1029/2023SW003729 (2024)

    doi:10.1029/2023sw003729 2024

  13. [13]

    The bruny island radio spectrometer.Publ

    Erickson, W. The bruny island radio spectrometer.Publ. Astron. Soc. Aust.14, 278–282, DOI: 10.1071/AS97278 (1997)

    work page doi:10.1071/as97278 1997

  14. [14]

    Gopalswamy, N.et al.Type II radio bursts and energetic solar eruptions.J. Geophys. Res. Space Phys.110, DOI: 10.1029/2005JA011158 (2005)

    doi:10.1029/2005ja011158 2005

  15. [15]

    Phys.289, 3059–3107, DOI: 10.1007/s11207-014-0524-8 (2014)

    Richardson, I.et al.>25 MeV proton events observed by the high energy telescopes on the STEREO A and B spacecraft and/or at Earth during the first seven years of the STEREO mission.Sol. Phys.289, 3059–3107, DOI: 10.1007/s11207-014-0524-8 (2014)

    doi:10.1007/s11207-014-0524-8 2014

  16. [16]

    Kasapis, S.et al.Interpretable machine learning to forecast SEP events for solar cycle 23.Space Weather.20, e2021SW002842, DOI: 10.1029/2021SW002842 (2022)

    work page doi:10.1029/2021sw002842 2022

  17. [17]

    F., Aydin, B., Martens, P

    Boubrahimi, S. F., Aydin, B., Martens, P. & Angryk, R. On the prediction of >100 MeV solar energetic particle events using GOES satellite data. In2017 IEEE International Conference on Big Data (Big Data), 2533–2542, DOI: 10.1109/BigData.2017.8258212 (IEEE, 2017)

    arXiv 2017

  18. [18]

    Phys.296, 107, DOI: 10.1007/s11207-021-01837-x (2021)

    Lavasa, E.et al.Assessing the predictability of solar energetic particles with the use of machine learning techniques.Sol. Phys.296, 107, DOI: 10.1007/s11207-021-01837-x (2021)

    doi:10.1007/s11207-021-01837-x 2021

  19. [19]

    A.et al.MEMPSEP-II

    Dayeh, M. A.et al.MEMPSEP-II. Forecasting the properties of solar energetic particle events using a multivariate ensemble approach.Space Weather.22, e2023SW003697, DOI: 10.1029/2023SW003697 (2024)

    work page doi:10.1029/2023sw003697 2024

  20. [20]

    A., Abdelkawy, A

    Ali, M. A., Abdelkawy, A. G., Shaltout, A. M. & Beheary, M. Forecasting solar energetic particles using multi- source data from solar flares, CMEs, and radio bursts with machine learning approaches.Sci. Reports15, 9546, DOI: 10.1038/s41598-025-92207-1 (2025). 10/12

    work page doi:10.1038/s41598-025-92207-1 2025

  21. [21]

    Space Res.72, 5161–5242, DOI: 10.1016/j.asr.2022.08.006 (2023)

    Whitman, K.et al.Review of solar energetic particle prediction models.Adv. Space Res.72, 5161–5242, DOI: 10.1016/j.asr.2022.08.006 (2023). 22.Yu, Y .et al.Solar energetic particle forecasting with multi-task deep learning: SEPNET (2026). 2512.12786

    Pith/arXiv arXiv 2022

  22. [23]

    G., Wright, P

    Bobra, M. G., Wright, P. J., Sun, X. & Turmon, M. J. SMARPs and SHARPs: Two solar cycles of active region data.The Astrophys. J. Suppl. Ser.256, 26, DOI: 10.3847/1538-4365/ac1f1d (2021)

    doi:10.3847/1538-4365/ac1f1d 2021

  23. [24]

    & Stegeman, L

    Whitman, K., Egeland, R., Allison, C., Quinn, P. & Stegeman, L. Validation of solar energetic particle forecasting models for space radiation operations with SPHINX and VIVID. Tech. Rep. NASA/TP-20260000463, National Aeronautics and Space Administration (2026)

    2026

  24. [25]

    ROC and AUC with a binary predictor: A potentially misleading metric.J

    Muschelli III, J. ROC and AUC with a binary predictor: A potentially misleading metric.J. classification37, 696–708, DOI: 10.1007/s00357-019-09345-1 (2020). 26.Lipton, Z. C., Elkan, C. & Narayanaswamy, B. Thresholding classifiers to maximize F1 score (2014). 1402.1892

    Pith/arXiv arXiv 2020

  25. [27]

    A., Brunelle, S

    Wehling, P., LaBudde, R. A., Brunelle, S. L. & Nelson, M. T. Probability of detection (pod) as a statistical model for the validation of qualitative methods.J. AOAC Int.94, 335–347 (2011)

    2011

  26. [28]

    Macmillan, N. A. & Kaplan, H. L. Detection theory analysis of group data: Estimating sensitivity from average hit and false-alarm rates.Psychol. bulletin98, 185 (1985)

    1985

  27. [29]

    & Keller, D

    Doswell, C., Davies-Jones, R. & Keller, D. L. On summary measures of skill in rare event forecasting based on contingency tables.Weather. forecasting5, 576–585 (1990)

    1990

  28. [30]

    D.et al.A comparison of flare forecasting methods

    Leka, K. D.et al.A comparison of flare forecasting methods. II. Benchmarks, metrics, and performance results for operational solar flare forecasting systems.The Astrophys. J. Suppl. Ser.243, 36, DOI: 10.3847/1538-4365/ab2e12 (2019). 31.Whitman, K. FetchSEP: CLEAR benchmark dataset version 1.0 (2025). Version CLEAR_Benchmark_v1.0, software

    doi:10.3847/1538-4365/ab2e12 2019

  29. [32]

    S.et al.The SunPy project: Open source development and status of the version 1.0 core package.The Astrophys

    Community, T. S.et al.The SunPy project: Open source development and status of the version 1.0 core package.The Astrophys. J.890, 68, DOI: 10.3847/1538-4357/ab4f7a (2020)

    doi:10.3847/1538-4357/ab4f7a 2020

  30. [33]

    Heliophysics events knowledgebase (HEK) (2026)

    Lockheed Martin Solar and Astrophysics Laboratory. Heliophysics events knowledgebase (HEK) (2026). Database, accessed 2026-03-16

    2026

  31. [34]

    G., Choudhary, N., Amezcua, A

    Glogowski, K., Bobra, M. G., Choudhary, N., Amezcua, A. B. & Mumford, S. J. drms: A Python package for accessing HMI and AIA data.J. Open Source Softw.4, 1614, DOI: 10.21105/joss.01614 (2019)

    doi:10.21105/joss.01614 2019

  32. [35]

    J.931, 163, DOI: 10.3847/1538-4357/ac64a6 (2022)

    Sun, Z.et al.Predicting solar flares using CNN and LSTM on two solar cycles of active region data.The Astrophys. J.931, 163, DOI: 10.3847/1538-4357/ac64a6 (2022)

    doi:10.3847/1538-4357/ac64a6 2022

  33. [36]

    A.et al.Time series of magnetic field parameters of merged MDI and HMI space-weather active region patches as potential tool for solar flare forecasting.The Astrophys

    Kosovich, P. A.et al.Time series of magnetic field parameters of merged MDI and HMI space-weather active region patches as potential tool for solar flare forecasting.The Astrophys. J.972, 169, DOI: 10.3847/1538-4357/ad60c3 (2024)

    work page doi:10.3847/1538-4357/ad60c3 2024

  34. [37]

    Space Weather Database of Notifications, Knowledge, Information (DONKI) (2026)

    NASA Community Coordinated Modeling Center. Space Weather Database of Notifications, Knowledge, Information (DONKI) (2026). Space weather event database. 38.Gopalswamy, N.et al.The SOHO LASCO CME catalog – version 2 (2025). 2407.04165. Acknowledgments The authors thank Ian Richardson (University of Maryland, NASA GSFC) and A. Steve Johnson (Leidos, NASA J...

    Pith/arXiv arXiv 2026