First-hand information for everyone
Main discovery ARENA (Active Region attENtion Aggregation) is an attention-based neural network that predicts whether an M‑ or X‑class (M+) solar flare will occur within 24 hours. It takes a variable number of active-region image cutouts as input and learns which region is most relevant to the prediction. (F1, F11)
How it works The problem is framed as multiple-instance learning (MIL). In MIL, the model sees a bag of instances and only a bag-level label. ARENA is trained under weak supervision using only flare/no‑flare labels for each observation. No labels identifying the flare‑producing region are provided during training. (F2)
Model architecture Each active-region (AR) cutout is encoded independently by a ResNet‑18 image encoder into a fixed-length embedding. The model aggregates the variable number of AR embeddings with a gated attention (ABMIL) mechanism. The attention function produces normalized weights α_i and the bag representation z is the attention‑weighted sum z = Σ_i α_i h_i. The attention weights give a localization signal that can be used at inference to identify the flare‑originating AR. (F3, F4)
Inputs and labels Inputs are multi‑channel Active Region Patches (AARPs) combining HMI magnetogram channels and AIA EUV channels (AIA 131, 171, 304 Å). Patches are resized to 224×224 pixels. Each timestamp is represented by 1–24 AARPs, with an average of about 7 AR cutouts per timestamp. A bag is labeled positive if at least one M‑ or X‑class flare occurs within the next 24 hours. (F5, F6)
Data and training The dataset uses solar observations from 2010–2024 sampled every two hours. Data from 2010–2016 were used for training and validation. Data from 2017–2024 were held out as an independent test set. Models were optimized with AdamW and cross‑entropy loss. Because M+ events are rare, negative‑instance undersampling was applied during training. Final reported hyperparameters include learning rate η = 1e‑7, weight decay 0.001, and batch size 32. (F7, F8)
Model size and evaluation The seven‑channel ResNet‑18 encoder has 11.2 million parameters. Instance embeddings have dimension M = 512. The gated attention uses attention dimension L = 128 (about 131,200 gated‑attention parameters). The classification head projects 512 → 256 (ReLU) → 1 and has 131,328 parameters. The primary forecasting metric is the Heidke Skill Score (HSS). Additional metrics include True Skill Statistic (TSS), precision, F1, TPR, TNR. AR source identification is quantified by ranking ARs by attention weight and reporting top‑k accuracy (Acc@k). (F9, F10)
Concrete results On the temporally held‑out 2017–2024 test set, ARENA achieves a median HSS of 0.45. (F11)
Despite never seeing flare‑origin annotations during training, ARENA assigns the highest attention weight to the annotated flare‑producing AR (Acc@1) in 87.1 ± 0.7% of evaluated cases. When excluding observations that contain only one AR, Acc@2 = 100.0 ± 0.0%. These localization results are evaluated on a subset with independent flare–AR associations. (F12, F15)
Comparisons On the same 2017 subset used for AR‑level comparison, ARENA (bag level, median across five temporal folds) reports: TSS = 0.79, HSS = 0.33, F1 = 0.37, Precision = 0.23, TPR = 0.92, TNR = 0.88. A per‑AR (instance‑level) classifier on that subset reports: TSS = 0.85, HSS = 0.17, F1 = 0.19, Precision = 0.10, TPR = 0.96, TNR = 0.90. These numbers show trade‑offs between instance‑level and bag‑level approaches on this evaluation subset. (F13)
Against other baselines, ARENA performs similarly to a low‑resolution full‑disk ResNet‑18 baseline across most metrics. It does not reach the reported performance of larger models such as P‑CNN and the fine‑tuned Surya model. ARENA exhibits the highest true positive rate among the compared models. Differences in training splits and model sizes complicate direct comparisons. (F14)
Significance and supported caveats ARENA demonstrates that attention‑based MIL can predict M+ flares from sets of AR images while delivering a localization signal. The attention weights can highlight the AR most likely associated with a predicted flare without requiring costly region‑level labels during training. (F1, F4)
Limitations to keep in mind. The AR source‑identification evaluation is restricted to a limited subset of data with independent flare–AR annotations available through 2017. Those annotations were used only for post‑hoc evaluation and not for training. (F15)
Class imbalance is a major dataset characteristic. Positive examples are rare. Metrics such as HSS, F1 and precision depend strongly on positive‑class prevalence. This complicates comparisons across different datasets and splits. (F16)
Augmenting the image‑only ARENA model with additional scalar inputs (SHARP magnetic summary parameters) or explicit AR spatial position and size did not yield consistent improvement and was omitted from the final model. (F17)
Overall, ARENA offers an interpretable, weakly supervised approach that jointly forecasts M+ flares and provides a localization signal pointing to likely source regions. Performance remains sensitive to dataset composition and to the choice of encoder; evaluation metrics also depend on class prevalence and dataset split.