ARENA: Active Region attENtion Aggregation for interpretable solar flare prediction

Luis Bauer; Annalena Kofler; Andrés Asensio Ramos; Aparna Venkataramanasastry; Valentín Martínez Pillet; Bernhard Schölkopf

astro-ph.SROct 23, 2026

ARENA: Active Region attENtion Aggregation for interpretable solar flare prediction

Authors: Luis Bauer, Annalena Kofler, Andrés Asensio Ramos, Aparna Venkataramanasastry, Valentín Martínez Pillet, Bernhard Schölkopf

Organizations: MPI for Intelligent Systems, Tübingen · University of Tübingen · MPI for Gravitational Physics, Potsdam · Instituto de Astrofísica de Canarias ·

Abstract

Predicting solar flares is a challenging space-weather forecasting problem with potentially significant impacts on Earth's upper atmosphere, including disruptions to radio communications. We propose Active Region attENtion Aggregation (ARENA), a neural network architecture that predicts solar flares from a variable number of active-region cutouts while learning to identify the region most relevant to the prediction. ARENA is trained using multiple-instance learning under weak supervision, requiring only a flare/no-flare label for the observation and no labels specifying the flare origin. The learned attention mechanism therefore provides a means of localizing the active region driving the prediction without explicit origin annotations. ARENA achieves a heidke skill score of 0.45 which is comparable to other flare prediction approaches. These results demonstrate that active-region-based aggregation can provide an efficient and interpretable approach to solar flare prediction. The code is available at https://github.com/LuisBauer22/ARENA.

Figures & tables

Figure 1: Overview of the ARENA architecture. Each AR across all channels is extracted, rescaled, and provided to a ResNet-18 which compresses the image into a token vector. For one time point, all AR tokens are concatenated and passed through an ABMIL aggregator, resulting in a single summary vector. This vector is processed by a classification head which predicts whether a flare is happening in the next $24\text{\,}\mathrm{h}$ . At inference, the attention weight can be employed to identify the flare-originating AR.

Panel of Figure 2 — Figure 2: (a) Heidke Skill Score (HSS) across the five temporal cross-validation folds, evaluated on the corresponding validation sets and on three test sets: the full 2017–2024 test period, the 2017 subset, and the 2017 subset containing flares with confirmed AR origins. (b) Density-normalized distribution of attention weights assigned to flare-producing and non-producing ARs for the 2017 subset with confirmed AR origins, using the model with the highest validation HSS among the five cross-validation models.

	TSS $\uparrow$	HSS $\uparrow$	F1 $\uparrow$	Precision $\uparrow$	TPR $\uparrow$	TNR $\uparrow$
Per AR (instance level)	0.85	0.17	0.19	0.10	0.96	0.90
ARENA (bag level)	0.79	0.33	0.37	0.23	0.92	0.88

Table 1: Comparison of per-AR instance flare level prediction vs bag level prediction on the same subset of 2017. We report the median metrics across the five validation folds.

Appendix figures & tables10 assets

Supplementary material from the paper’s appendix.

Appendix

Figure 5: Number of ARs per timestamp across different datasets.

Figure 6: Number of flares per year and dataset split. We use the data from 2010-2016 for training and validation. The data from 2017-2024 is used as the holdout test-set. On a subset of the 2017 test data (blue) we evaluate the attention weights for each region as this overlaps with our flare catalog with AR labels [ 1 ] .

Figure 7: Confusion matrices for the different temporally held-out test subsets. The model correctly identifies a majority of both positive and negative observations, with consistent performances across each evaluation set. This can be observed when comparing TP and FN as well as TN and FP values, the model exhibits good predictive performance. This is also the case for the ratio of FN and TN values. However, comparing TPs and FPs, we observe that the model predicts a large number of FP values, especially in the 2017 test sets. The background colors are normalized per row of each matrix.

Dataset	TSS $\uparrow$	HSS $\uparrow$	F1 $\uparrow$	Precision $\uparrow$	TPR $\uparrow$	TNR $\uparrow$
Validation	$0.48^{+0.05}_{-0.08}$	$0.42^{+0.00}_{-0.09}$	$0.54^{+0.05}_{-0.01}$	$0.46^{+0.05}_{-0.04}$	$0.70^{+0.04}_{-0.05}$	$0.77^{+0.02}_{-0.02}$
Test 2017–2024	$0.61^{+0.01}_{-0.01}$	$0.45^{+0.01}_{-0.01}$	$0.62^{+0.01}_{-0.01}$	$0.47^{+0.01}_{-0.01}$	$0.94^{+0.00}_{-0.00}$	$0.67^{+0.01}_{-0.01}$
Test 2017	$0.69^{+0.00}_{-0.00}$	$0.36^{+0.01}_{-0.03}$	$0.41^{+0.01}_{-0.03}$	$0.28^{+0.00}_{-0.03}$	$0.82^{+0.00}_{-0.01}$	$0.88^{+0.00}_{-0.02}$
AR subset 2017	$0.79^{+0.01}_{-0.02}$	$0.33^{+0.01}_{-0.04}$	$0.37^{+0.01}_{-0.03}$	$0.23^{+0.00}_{-0.03}$	$0.92^{+0.00}_{-0.00}$	$0.88^{+0.00}_{-0.02}$

Appendix

Table 2: Additional performance metrics on the different evaluation sets in the format

median^{Q75}_{Q25}

	Training set	TSS $\uparrow$	HSS $\uparrow$	F1 $\uparrow$	Precision $\uparrow$	TPR $\uparrow$	TNR $\uparrow$
P-CNN (ensemble) [ 10 ]	2010–2019	0.58	0.42	0.53	0.39	0.82	0.76
Full disk (ensemble)	2010–2017	0.52	0.34	0.49	0.34	0.81	0.71
Fine-tuned Surya [ 24 ]	2010–2017	0.48	0.46	0.55	0.54	0.57	0.90
ARENA (ensemble)	2010–2017	0.51	0.30	0.46	0.31	0.87	0.64

Appendix

Table 3: Baseline comparison of ARENA on the January 2020–April 2023 subset of the test period, corresponding to the evaluation period reported for P-CNN [ 10 ] . ARENA and the full-disk ResNet, were implemented and trained by the authors; P-CNN results are taken from the original paper [ 10 ] . Pre-processing and training protocols differ between methods and therefore the comparison should not be interpreted as a fully controlled benchmark. Detailed experimental configurations are provided in Appendix D .

Panel of Figure 8 — Figure 8: (a) TSS vs. positive-class prevalence: The evaluation subsets with lower positive-class prevalence tend to exhibit higher TSS, with the 2017 and 2017 AR-annotated subsets reaching values of approximately 0.69–0.80, compared with 0.40–0.62 for the other subsets. (b) HSS vs. positive-class prevalence: HSS shows a weaker tendency in the opposite direction, with values of approximately 0.29–0.42 for the highly imbalanced 2017 subsets and approximately 0.45 for the full test set. The validation subsets show no clear monotonic relationship.

Panel of Figure 9 — Figure 9: Temporal distribution of the four binary classification outcomes for the 2017 test subset. True positives and false positives are shown among the predicted-positive observations, while true negatives and false negatives constitute the predicted-negative observations. False positives frequently occur in temporal proximity to true positives.

Figure 10: Temporal evolution of ARENA’s predicted M+ flare probability (top) and attention weights assigned to individual active regions (bottom) for July 6th–19th, 2017. The orange shading indicates observations labeled positive according to our flare catalog because an M+ flare occurs within the subsequent 24-hour prediction window. Flares are represented by the black vertical dashed lines. The same active region receives high attention weights

\alpha_{i}

across multiple consecutive observations surrounding the two flare events, while the predicted flare probability remains elevated over an extended period. As soon as this region disappears on July 18th, the output probability of the model drops. The yellow line corresponds to SHARP 7075 which is actually causing the two flares according to [ 1 ] .

Attention Mechanisms

Attention mechanisms let neural networks weight information from other tokens or regions according to relevance, enabling context-dependent processing.

See all 517 papers

Prediction of Major Solar Flares Using Interpretable Class-dependent Reward Framework with Active Region Magnetograms and Domain Knowledge
Feb 18, 2026Zixian Wu, Xuebao Li, Yanfang Zheng +14Solar Physics Cost-Sensitive Learning
School of Computer Science, Jiangsu University of Science and Technology, Zhenjiang 212100, People′s Republic of China · State Key Laboratory of Space Weather, Chinese Academy of Sciences, Beijing 100190, People′s Republic of China · School of software, Southeast University, Nanjing, People′s Republic of China +4
Predicting Associations between Solar Flares and Coronal Mass Ejections Using SDO/HMI Magnetograms and a Hybrid Neural Network
Apr 11, 2026Jialiang Li, Vasyl Yurchyshyn, Jason T. L. Wang +6Solar Physics
Institute for Space Weather Sciences, New Jersey Institute of Technology, University Heights, Newark, NJ 07102, USA · Department of Computer Science, New Jersey Institute of Technology, University Heights, Newark, NJ 07102, USA · Big Bear Solar Observatory, New Jersey Institute of Technology, 40386 North Shore Lane, Big Bear City, CA 92314, USA +4
A Deep Learning Framework for Predicting Solar EUV Irradiance During Significant Flares
Jul 21, 2026Sathvik Soman, Jason T. L. Wang, Haimin Wang +1Solar Irradiance Forecasting Time Series Forecasting
New Jersey Institute of Technology, Newark, NJ, USA · Sam Houston State University, Huntsville, TX, USA

Feb 18, 2026J/K move · Enter open · S save

Prediction of Major Solar Flares Using Interpretable Class-dependent Reward Framework with Active Region Magnetograms and Domain Knowledge

Zixian Wu, Xuebao Li, Yanfang Zheng, Rui Wang+13

School of Computer Science, Jiangsu University of Science and Technology, Zhenjiang 212100, People′s Republic of China · State Key Laboratory of Space Weather, Chinese Academy of Sciences, Beijing 100190, People′s Republic of China · School of software, Southeast University, Nanjing, People′s Republic of China+4

Solar Physics Cost-Sensitive Learning

Open paper AskPDF arXiv

In this work, we develop, for the first time, a supervised classification framework with class-dependent rewards (CDR) to predict

\geq

MM flares within 24 hr. We construct multiple datasets, covering knowledge-informed features and line-of sight (LOS) magnetograms. We also apply three deep learning models (CNN, CNN-BiLSTM, and Transformer) and three CDR counterparts (CDR-CNN, CDR-CNN-BiLSTM, and CDR-Transformer). First, we analyze the importance of LOS magnetic field parameters with the Transformer, then compare its performance using LOS-only, vector-only, and combined magnetic field parameters. Second, we compare flare prediction performance based on CDR models versus deep learning counterparts. Third, we perform sensitivity analysis on reward engineering for CDR models. Fourth, we use the SHAP method for model interpretability. Finally, we conduct performance comparison between our models and NASA/CCMC. The main findings are: (1)Among LOS feature combinations, R_VALUE and AREA_ACR consistently yield the best results. (2)Transformer achieves better performance with combined LOS and vector magnetic field data than with either alone. (3)Models using knowledge-informed features outperform those using magnetograms. (4)While CNN and CNN-BiLSTM outperform their CDR counterparts on magnetograms, CDR-Transformer is slightly superior to its deep learning counterpart when using knowledge-informed features. Among all models, CDR-Transformer achieves the best performance. (5)The predictive performance of the CDR models is not overly sensitive to the reward choices.(6)Through SHAP analysis, the CDR model tends to regard TOTUSJH as more important, while the Transformer tends to prioritize R_VALUE more.(7)Under identical prediction time and active region (AR) number, the CDR-Transformer shows superior predictive capabilities compared to NASA/CCMC.

Abstract

Attention Mechanisms

Explore similar work

Prediction of Major Solar Flares Using Interpretable Class-dependent Reward Framework with Active Region Magnetograms and Domain Knowledge

Predicting Associations between Solar Flares and Coronal Mass Ejections Using SDO/HMI Magnetograms and a Hybrid Neural Network

A Deep Learning Framework for Predicting Solar EUV Irradiance During Significant Flares