Source-linked AI summary
ARENA: Active Region attENtion Aggregation for interpretable solar flare prediction
Luis Bauer, Annalena Kofler, Andrés Asensio Ramos, Aparna Venkataramanasastry, Valentín Martínez Pillet, Bernhard Schölkopf
TL;DR
Solar flares are global events with localized active-region origins, making forecasting and source identification distinct challenges. ARENA applies multiple-instance learning to variable sets of active-region cutouts using only observation-level labels. It reports a median HSS of 0.45 on the held-out 2017–2024 test set and identifies the flare-producing region without origin annotations during training.
Problem
Flare forecasting must predict whether an event occurs within 24 hours, while source identification requires linking it to a localized active region.
Source: §1 · p. 1Method
ARENA encodes variable sets of active-region cutouts and uses attention-based multiple-instance learning trained on observation-level flare labels.
Source: §1 · p. 2Results
Takeaways & Limitations
ARENA provides an active-region importance signal while maintaining forecasting performance comparable to a model trained on flare–active-region labels.
Source: §5 · p. 5The flare-producing region is assumed to be among the active regions present at observation time.
Source: §2 · p. 3
Abstract
from arXiv · showhide
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.
1 Introduction
Solar flares are global forecasting targets with localized active-region origins, making both event prediction and source identification relevant. ARENA uses multiple-instance learning to predict from variable sets of active regions and learn a localization signal without flare-origin labels.
1 Introduction
- Flare forecasting asks whether an event occurs within 24 hours, while identifying its origin requires associating the flare with a source active region.Flare radiation can affect radio communication and space-based technologies.Source: §1 · p. 1
- ARENA combines active-region instances as variable-sized sets, representing the solar disk with physically meaningful regions rather than full-disk images or isolated regions.The introduction frames the set representation as a natural way to combine the active regions present in an observation.Source: §1 · p. 2
- Origin annotations require expert event-to-region associations and are available only for limited datasets, whereas observation-level flare labels are readily available.The paper frames this mismatch as a multiple-instance learning problem: observations are bags, active-region cutouts are instances, and labels apply to bags.Source: §1 · p. 2
- ARENA learns from observation-level flare labels, and its attention weights can identify the flare-associated active region despite no origin annotations during training.The authors report comparable flare-forecasting performance while providing an intrinsic active-region localization signal.Source: §1 · p. 2
2 ARENA: Active Region attENtion Aggregation
ARENA treats each solar observation as a bag of active-region image cutouts and predicts a 24-hour flare label from their aggregated representations. Training uses only bag-level labels, with the flare-producing region unannotated.
2 ARENA: Active Region attENtion Aggregation
- ARENA represents an observation as a bag of active-region cutouts whose instance count varies; observations with no active regions are excluded.The active-region representation focuses on physically relevant regions and reduces memory and disk requirements relative to processing full-disk images.Source: §2 · p. 2
- Each bag’s binary label marks whether at least one M- or X-class flare occurs in the following 24 hours.For positive bags, the flare-producing region is assumed to be among the active regions present at the observation time.Source: §2 · p. 3
- Training provides only the bag-level label, so the model must learn relevant active regions without flare-origin annotations.This is weak supervision at the active-region level and motivates attention aggregation over a variable number of instances.Source: §2 · p. 3
- ARENA encodes each active-region cutout with ResNet-18, aggregates embeddings using ABMIL attention, then maps the bag representation to a 24-hour flare probability.The learned attention-weighted sum allows the model to select relevant representations and features for the bag-level prediction.Source: §2 · p. 3
3 Experimental Setup
The study uses two-hourly solar observations, a temporally held-out test period, and five temporal validation folds. Evaluation covers both flare forecasting and localization of flare-producing active regions.
3 Experimental Setup
- Observations from 2010–2024 are sampled every two hours; 2010–2016 support training and validation, while 2017–2024 form the held-out test set.Five temporal folds use two validation months per year and a one-month buffer between training and validation periods.Source: §3 · p. 3
- HSS is the primary flare-forecasting metric across validation and held-out test sets, with comparison to an instance-level classifier trained using explicit flare–active-region annotations.Additional classification metrics characterize differences between direct active-region-level and bag-level prediction.Source: §3 · p. 3
- Source identification ranks active regions in positive bags by attention weight and measures whether the annotated flare-producing region appears among the top k.The evaluation also compares attention-weight distributions for flare-producing and non-producing regions; Figure 2 covers five folds and three test sets.Source: §3 · pp. 3–4, Figure 2 · p. 4
4 Results
On the temporally held-out 2017–2024 test set, ARENA achieves a median HSS of 0.45 and performs comparably to an AR-level supervised baseline. Its attention weights also identify flare-producing regions with high accuracy and are systematically higher for those regions.
Flare forecasting on future data
- Median test HSS is 0.45, slightly above the validation median of 0.42, and ARENA performs comparably to the AR-level supervised baseline.The test set covers 2017–2024; Table 1 compares per-AR instance and bag-level prediction on the same 2017 subset, reporting median metrics across five validation folds.Source: §4 · p. 4, Table 1 · p. 4
AR source identification
- Excluding observations containing only one AR, ARENA achieves source-identification Acc@1 of 87.1 ± 0.7% and Acc@2 of 100.0 ± 0.0%.Source: §4 · p. 4
- Flare-producing ARs receive systematically higher attention weights than non-producing ARs across the five models.The plotted distributions are for the best run, which had validation HSS=0.47.Source: §4 · p. 5
5 Conclusion
ARENA uses bag-level flare labels and AR cutouts to forecast flares, achieving a median HSS of 0.45 on the temporally held-out 2017–2024 test set. Its attention identifies the flare-producing AR in 87.1 ± 0.7% of cases despite training without flare-origin annotations.
Conclusion
- ARENA is an attention-based multiple-instance learning architecture for interpretable solar flare forecasting.It uses bag-level flare labels and AR cutouts, reaches a median HSS of 0.45 on the 2017–2024 held-out test set, and assigns the highest attention weight to the flare-producing AR in 87.1 ± 0.7% of cases.Source: §5 · p. 5
A Data set and preprocessing
The dataset represents each observation as a variable-sized set of active-region cutouts, combining magnetic-field and coronal-morphology information. Two-hour sampling, channel-specific preprocessing, future-period testing, and limited origin annotations define the dataset and its evaluation scope.
Active region image catalog
- SHARPs provides automatically identified and tracked active-region patches from HMI observations, focusing inputs on flare-relevant regions rather than quiet-Sun areas.Using AR cutouts reduces memory, computational, and storage requirements; the catalog is derived from photospheric magnetic-field observations.Source: A Data set and preprocessing · p. 8
Multi-channel active-region dataset
- Co-aligned AIA channels enter the network separately, allowing it to use channel-specific and cross-channel morphological differences without explicitly representing temperature.Source: A Data set and preprocessing · p. 8
- AARPs combine photospheric magnetic-field information with AIA observations that provide coronal and transition-region morphology.The 131, 171, and 304 Å AIA channels sample plasma across a range of characteristic formation temperatures; channel intensities are not reconstructed into temperature maps.Source: A Data set and preprocessing · p. 8
- At two-hour cadence, each timestamp contains 1–24 AARPs, averaging approximately seven active-region cutouts.The cutouts are aligned across HMI and AIA observations from the Surya dataset.Source: A Data set and preprocessing · p. 8
Data normalization and spatial preprocessing
- EUV intensities receive log(1 + x) transformation and channel-wise standardization, while magnetograms are clipped to [−300, 300] Gauss and rescaled to [−1, 1].All variable-sized patches are resized to 224 × 224 pixels; adding original position and size information did not significantly improve forecasting performance.Source: A Data set and preprocessing · p. 10
Flare labels
- Observations are labeled positive when an M- or X-class flare occurs within the following 24 hours, using GOES XRS flare records.Training and validation use 2010–2016, while 2017–2024 is held out; five temporal validation folds use adjacent-month buffers to reduce leakage from temporal correlations.Source: A Data set and preprocessing · pp. 10–11
Flare origin
- SWAN flare–AR associations cover 2010–2017, and matching to GOES XRS within ±1 hour covers approximately 65% of flares in that period.For source identification evaluation and encoder comparison, timestamps with a subsequent flare but no SWAN source annotation are discarded.Source: A Data set and preprocessing · p. 11
- Because reliable flare–AR associations are available only for a limited subset, source-region evaluation is restricted to annotated M+ flares in 2017.The source annotations are reserved for evaluation, not training; incomplete SWAN coverage motivates restricting that evaluation subset.Source: A Data set and preprocessing · p. 11
B Architecture details
ARENA encodes each active-region cutout and aggregates the embeddings with normalized attention into a bag representation for flare classification. Its gated attention is permutation-invariant and accommodates bags with different numbers of regions.
B Architecture details
- Normalized attention weights combine active-region embeddings into an aggregation vector whose magnitude is independent of the number of regions.The vector is passed to a two-layer classifier that maps the 512-dimensional representation through 256 units to one classification logit.Source: B Architecture details · pp. 11–12
- ARENA uses a gated attention mechanism that multiplies two transformed versions of each embedding before projecting to a scalar score.The encoder is a seven-channel ResNet-18 with 11.2 million parameters; the attention module uses M = 512 and L = 128.Source: B Architecture details · pp. 11–12
- The normalized attention-weighted aggregation is invariant to the ordering of active regions and supports observations with different numbers of regions.Softmax normalization makes the aggregation-vector magnitude independent of bag sequence length.Source: B Architecture details · pp. 11–12
C Training details
Negative-instance undersampling produced the best validation performance among the tested imbalance strategies and was used in all reported experiments.
C Training details
- Negative-instance undersampling yielded the best validation performance and was used for all reported experiments.Undersampling was applied only to training data; validation and test sets retained their original class distributions.Source: C Training details · p. 12
- Without undersampling, models tended to predict nearly all observations as negative; weighted sampling and class-weighted loss reduced this tendency without consistent improvement over undersampling.The final configuration used a learning rate of 10^-7, weight decay of 0.001, and batch size of 32; higher learning rates degraded validation performance.Source: C Training details · p. 12
D Evaluation
Evaluation covers flare forecasting and source-region identification, with HSS as the primary forecasting metric. Comparisons use explicit caveats for differences in class composition, data splits, and preprocessing.
D Evaluation
- HSS is the primary metric for forecasting evaluation, with TSS, precision, and F1 also reported from confusion-matrix counts.A 0.5 threshold is applied to model output probabilities unless stated otherwise.Source: D Evaluation · pp. 12–13
- ARENA and the instance-level classifier share an encoder and temporal splits, but their HSS, F1, and precision comparisons are not fully controlled because positive prevalence differs.Positive examples are 1.2% of instance-level data versus 3.86% of ARENA bags; TSS, TPR, and TNR offer an indication of whether weakly supervised aggregation limits performance.Source: D Evaluation · p. 13
- Source-region identification ranks active regions by normalized attention on the 2017 test subset with known SWAN source associations, used only for post-hoc evaluation.Annotations are excluded from training, model selection, and attention computation.Source: D Evaluation · p. 13
- The low-resolution full-disk model is the most relevant baseline; Surya and P-CNN are not directly comparable because of different data splits and preprocessing.The full-disk model uses the same seven channels, ResNet-18 encoder, and classification head as ARENA; Surya evaluation avoids observations used in pretraining, while P-CNN results are taken from its publication.Source: D Evaluation · p. 14
- Additional analyses examine confusion matrices, metrics across validation and temporally held-out subsets, class prevalence, and false positives on a test set.The inspected validation fold had HSS = 0.47.Source: D Evaluation · p. 14
E Additional results
Additional evaluations compare ARENA with other models, test optional input features, and examine how class prevalence relates to performance. They also investigate false positives and the limits of interpreting results across differently balanced subsets.
Flare prediction and cross-validation
- Across evaluation subsets, ARENA maintains relatively high true-positive and true-negative rates, but its TP-to-FP ratio falls on the 2017 tests.The test confusion matrices show many false positives, especially in the 2017 sets.Source: E Additional results · p. 15, Figure 7 · p. 16
Extra scalar parameters
- Adding SHARP magnetic summaries or AR position and size does not consistently improve on image-only ARENA.Source: E Additional results · p. 15
Baseline comparison
- ARENA performs similarly to the full-disk baseline across most metrics but does not match P-CNN or Surya, while achieving the highest TPR.The comparison is not fully controlled because training splits and protocols differ; Surya has roughly 20× more parameters.Source: E Additional results · p. 15, Table 3 · p. 16
Metric dependence on class imbalance
- The 2017 subsets show higher TSS but lower HSS than the full test set, so apparent performance depends on the metric and evaluation-set prevalence.The high TSS for the 2017 AR subset warrants caution because it is estimated from comparatively few positive observations.Source: E Additional results · pp. 15–17, Figure 8 · p. 17
False-positive analysis
- About three times as many false positives as true positives occur in the 2017 subset, and false positives often cluster near true positives.The authors suggest the model may track elevated AR flare propensity over time while the 24-hour binary target imposes a sharper temporal boundary.Source: E Additional results · p. 17
F Related work
Prior image-based flare forecasting uses full-disk images or AR images with flare labels, while one related MIL approach uses solar spectra. The supplied figure passages also describe temporally clustered false positives and attention concentrated on the AR associated with two flares.
F Related work
- Image-based forecasting approaches use full-disk images or AR images with flare labels, while Huwyler and Melchior [31] apply MIL to solar spectra.The passage notes that most flares originate from ARs [1].Source: F Related work · p. 17
- False positives in the 2017 test subset frequently occur near true positives in time.Source: Figure 9 · p. 18
- Across two flare events, the same AR receives high attention over consecutive observations, and predicted probability drops when that region disappears.The figure identifies SHARP 7075 as the region causing both flares according to [9].Source: Figure 10 · p. 19