Deep SWM: Solar Flare Forecast Model

• Deep SWM is a deep state-space model for long-range solar flare forecasting using multi-wavelength imagery and sparse masked autoencoder pretraining.
• It integrates dual branches—one for spatial encoding via the Solar Spatial Encoder and another for long-term dependencies via LT-SSM—to capture complex spatio-temporal patterns.
• Evaluated on FlareBench using metrics like GMGS, TSS, and BSS, Deep SWM demonstrates superior reliability and performance over previous models and expert forecasts.

Searching arXiv for the specified Deep SWM paper and the earlier similarly named space weather model to ground the article in the cited literature. {"query": "arXiv (Nagashima et al., 11 Aug 2025) Deep Space Weather Model Long-Range Solar Flare Prediction from Multi-Wavelength Images", "max_results": 5} {"query": "(Nagashima et al., 11 Aug 2025)", "max_results": 10} Deep Space Weather Model (Deep SWM) is a solar flare prediction model introduced for long-range forecasting from multi-wavelength solar imagery. It is designed around multiple deep state space models that jointly handle ten-channel solar images and long-range spatio-temporal dependencies, and it is coupled with a sparse masked autoencoder pretraining strategy and a public benchmark, FlareBench, covering a full 11-year solar activity cycle (Nagashima et al., 11 Aug 2025). In this formulation, the central problem is multi-class prediction of the maximum GOES X-ray class in the next 24 hours from recent solar observations, while addressing two limitations identified in prior work: heuristic physical features often lack representation learning from solar images, and end-to-end learning approaches struggle to model long-range temporal dependencies.

1. Problem formulation and predictive target

Deep SWM takes as input a sequence of $k$ observations,

$x = (V_{t-k+1}, \ldots, V_t),$

where each $V_t \in \mathbb{R}^{C \times H \times W}$ is a $C=10$-channel solar image consisting of $1$ HMI magnetogram and $9$ AIA wavelengths (Nagashima et al., 11 Aug 2025). In FlareBench, the operational setting uses $k=4$ frames at $1$-hour cadence. The prediction target is the maximum GOES X-ray class in the next $24\,\mathrm{h}$, with four classes: $X$, $x = (V_{t-k+1}, \ldots, V_t),$0, $x = (V_{t-k+1}, \ldots, V_t),$1, or $x = (V_{t-k+1}, \ldots, V_t),$2.

The benchmark spans a continuous 11-year cycle from June 2011 to November 2022 and contains $x = (V_{t-k+1}, \ldots, V_t),$3 samples after cleaning. The class counts are highly imbalanced: $x = (V_{t-k+1}, \ldots, V_t),$4; $x = (V_{t-k+1}, \ldots, V_t),$5; $x = (V_{t-k+1}, \ldots, V_t),$6; $x = (V_{t-k+1}, \ldots, V_t),$7. Time-series cross-validation uses $x = (V_{t-k+1}, \ldots, V_t),$8 folds, each with distinct solar cycle phases in test. This setup makes the model a sequence classifier over multi-channel image streams rather than a single-image detector or a univariate flare-threshold forecaster.

A plausible implication is that the formulation explicitly separates short-horizon observation windows from longer latent histories: the supervised input remains $x = (V_{t-k+1}, \ldots, V_t),$9 recent frames, but the model also incorporates a longer pretraining history $V_t \in \mathbb{R}^{C \times H \times W}$0 in its temporal branch.

2. Architectural design and deep state-space modeling

The architecture has two parallel branches followed by fusion and prediction (Nagashima et al., 11 Aug 2025). The first branch is the Solar Spatial Encoder (SSE), which produces $V_t \in \mathbb{R}^{C \times H \times W}$1. It captures multi-scale spatio-temporal features via 3D conv downsampling, followed by a Depth-wise Channel Selective Module (DCSM), then a Spatio-Temporal SSM (ST-SSM) implemented with S5 blocks. The branch is hierarchical: it iterates $V_t \in \mathbb{R}^{C \times H \times W}$2 stages and then flattens to sequence length $V_t \in \mathbb{R}^{C \times H \times W}$3 and feature dimension $V_t \in \mathbb{R}^{C \times H \times W}$4.

The second branch combines Sparse-MAE pretraining with a Long-range Temporal SSM (LT-SSM), producing $V_t \in \mathbb{R}^{C \times H \times W}$5. Sparse masked autoencoder pretraining operates on a much longer history $V_t \in \mathbb{R}^{C \times H \times W}$6 and yields per-frame features

$V_t \in \mathbb{R}^{C \times H \times W}$7

The LT-SSM then applies $V_t \in \mathbb{R}^{C \times H \times W}$8 stacked SSMBlock layers to $V_t \in \mathbb{R}^{C \times H \times W}$9 to capture dependencies over $C=10$0 one solar rotation.

Fusion proceeds by concatenating along the sequence dimension, passing the result through one more SSMBlock and a feed-forward network:

$C=10$1

This yields class probabilities over $C=10$2.

Each SSMBlock is abstracted as a pair of functions $C=10$3 and $C=10$4. The hidden transition is

$C=10$5

and in Deep SWM, $C=10$6 is implemented by an SSMBlock that combines a Multi-Input Multi-Output linear state-space layer $C=10$7, based on S5, and a small MLP. For input $C=10$8,

$C=10$9

$1$0

The observation function is

$1$1

where $1$2 is a $1$3 conv or small FFN that maps the final latent sequence to per-time-step logits, followed by softmax.

The ST-SSM and LT-SSM divide the spatio-temporal modeling problem into two regimes. After channel weighting by DCSM, ST-SSM flattens the feature map $1$4 over $1$5 positions and feeds it into an SSMBlock, enabling each spatial location’s history to be processed by the same long-range SSM. LT-SSM instead operates on per-frame feature vectors and stacks multiple SSMBlocks across the full history. The use of S5 is motivated in the paper by a time-invariant MIMO SSM layer with efficient $1$6 complexity in sequence length, diagonalization of $1$7 via eigen-decomposition, Zero-Order Hold discretization, and HiPPO initialization for robust memorization of long sequences.

3. Sparse masked autoencoder pretraining

Deep SWM includes a sparse masked autoencoder as a pretraining mechanism for the image encoder (Nagashima et al., 11 Aug 2025). Its defining feature is a two-phase masking strategy intended to preserve crucial regions such as sunspots while compressing spatial information.

Given an image $1$8, the first phase is spatial-level masking. Patch-wise standard deviation is computed, the top $1$9 high-variance patches are selected, and these “important” patches are masked with a low ratio $9$0, while the remaining patches are masked with a higher ratio $9$1, with $9$2. The second phase is feature-level masking: after mapping visible patches through the encoder, the encoded features are further randomly masked with ratio $9$3.

The reconstruction objective is

$9$4

where $9$5 is the set of masked pixel positions.

The stated downstream effect is twofold. First, the encoder is encouraged to learn detailed representations of sparse yet crucial regions, especially sunspots. Second, initializing ST-SSM and LT-SSM downstream layers with these weights leads to faster convergence and higher predictive accuracy, as shown in the ablation study. This suggests that the pretraining stage is not merely a generic compression step; it is explicitly tuned to sparse active-region structure.

4. Data benchmark, optimization, and inference pipeline

FlareBench is the evaluation benchmark constructed to validate Deep SWM over a full solar cycle (Nagashima et al., 11 Aug 2025). The benchmark uses continuous data from June 2011 to November 2022, time-series cross-validation with three folds, and data augmentation consisting of random rotation, scaling, brightness/contrast jitter, Gaussian blur, and channel-specific noise.

Training follows a two-stage procedure. In the first stage, the full model is trained on the original data for $9$6 epochs. In the second stage, classifier re-training is performed for $9$7 epochs to mitigate class imbalance via CRT sampling. The optimizer is AdamW with $9$8, $9$9, batch size $k=4$0, learning rate $k=4$1 in both stages, and weight decay $k=4$2.

Inference is specified as a four-step pipeline. First, a new $k=4$3-frame sequence is preprocessed and standardized, then fed into the Sparse-MAE encoder without masking to obtain $k=4$4 frames. Second, the SSE branch applies downsampling, DCSM, and ST-SSM, then flattens to $k=4$5. Third, the LT-SSM branch stacks SSMBlocks on $k=4$6 to obtain $k=4$7. Fourth, fusion is performed by

$k=4$8

The benchmark and training protocol are integral to the reported claims. The paper attributes validation of performance and reliability not only to the network design, but also to evaluation over distinct solar cycle phases and explicit treatment of class imbalance.

5. Evaluation metrics, quantitative results, and reliability

Deep SWM is evaluated with three metrics (Nagashima et al., 11 Aug 2025). For binary discrimination of $k=4$9 versus $1$0 events, the True Skill Statistic is

$1$1

Probabilistic calibration for large flares is measured by $1$2, with

$1$3

$1$4

The multi-class metric is the Gandin–Murphy–Gerrity Score,

$1$5

where $1$6 is the empirical contingency table and $1$7 is the symmetric scoring matrix defined in Eqs. (16)–(19) of the paper.

Method	Period	Selected results
Flare Transformer (full)	2014–2017	GMGS $1$8; $1$9 $24\,\mathrm{h}$0; $24\,\mathrm{h}$1 $24\,\mathrm{h}$2
CNN-LSTM	2019–2022	GMGS $24\,\mathrm{h}$3; $24\,\mathrm{h}$4 $24\,\mathrm{h}$5; $24\,\mathrm{h}$6 $24\,\mathrm{h}$7
Ours (Deep SWM)	2019–2022	GMGS $24\,\mathrm{h}$8; $24\,\mathrm{h}$9 $X$0; $X$1 $X$2
Human experts	2000–2015	GMGS $X$3; $X$4 $X$5; $X$6 $X$7

The paper states that the method outperformed baseline methods and even human expert performance on standard metrics in terms of performance and reliability. For reliability analysis, $X$8 is interpreted as measuring probabilistic calibration for large flares, and Deep SWM’s $X$9 is described as far exceeding baselines. Calibration curves, not shown in the summary, are reported to indicate near-diagonal reliability for $x = (V_{t-k+1}, \ldots, V_t),$00 forecasts. The uncertainty discussion in the paper emphasizes that softmax output probabilities align with observed frequencies, improving operational decision-support.

A common misconception is to treat higher classification accuracy as sufficient for operational flare forecasting. The reported evaluation instead separates categorical discrimination, probabilistic calibration, and multi-class scoring through TSS, BSS, and GMGS, indicating that reliability is a first-class objective rather than a secondary diagnostic.

6. Ablations, interpretive findings, and naming ambiguity

The ablation study isolates the contribution of the principal modules (Nagashima et al., 11 Aug 2025). Removing LT-SSM reduces GMGS to $x = (V_{t-k+1}, \ldots, V_t),$01. Removing DCSM yields GMGS $x = (V_{t-k+1}, \ldots, V_t),$02. Removing ST-SSM yields GMGS $x = (V_{t-k+1}, \ldots, V_t),$03. Removing the entire SSE yields GMGS $x = (V_{t-k+1}, \ldots, V_t),$04. The full model reaches GMGS $x = (V_{t-k+1}, \ldots, V_t),$05. The paper’s direct conclusion is that every module contributes, with LT-SSM having the largest single impact.

Pretraining ablations further distinguish the sparse masking strategy from conventional masked autoencoding. MAE$x = (V_{t-k+1}, \ldots, V_t),$06 yields GMGS $x = (V_{t-k+1}, \ldots, V_t),$07 with $x = (V_{t-k+1}, \ldots, V_t),$08. MAE$x = (V_{t-k+1}, \ldots, V_t),$09 yields GMGS $x = (V_{t-k+1}, \ldots, V_t),$10 with $x = (V_{t-k+1}, \ldots, V_t),$11. Sparse MAE yields GMGS $x = (V_{t-k+1}, \ldots, V_t),$12 with $x = (V_{t-k+1}, \ldots, V_t),$13. The stated conclusion is that two-phase masking is crucial for reconstructing sunspot features.

The paper also provides several interpretive insights: DCSM’s channel weighting highlights the most flare-relevant wavelengths; ST-SSM captures fine-grained spatio-temporal interactions within each stage; LT-SSM extends memory well beyond $x = (V_{t-k+1}, \ldots, V_t),$14 hours and is essential for regions evolving over solar rotation periods; and Sparse MAE pretraining seeds the encoder with strong representations of sparse active regions, yielding dramatic improvements in classification accuracy and reliability. These points remain interpretations reported by the authors, but they are consistent with the ablation pattern.

The acronym “Deep SWM” also has an earlier, distinct use in Benson et al. for “Simultaneous Multivariate Forecast of Space Weather Indices using Deep Neural Network Ensembles” (Benson et al., 2021). In that work, the model is a 3-layer stacked LSTM with hidden size $x = (V_{t-k+1}, \ldots, V_t),$15, trained to forecast $x = (V_{t-k+1}, \ldots, V_t),$16 daily-cadence indices over a $x = (V_{t-k+1}, \ldots, V_t),$17-day horizon from $x = (V_{t-k+1}, \ldots, V_t),$18-day input windows, with optional $x = (V_{t-k+1}, \ldots, V_t),$19-dimensional solar-image embeddings. That earlier model concerns simultaneous multivariate forecasting of space weather indices rather than multi-class flare prediction from ten-channel solar images. The shared acronym can therefore be misleading: the 2025 Deep SWM denotes a deep state-space architecture for flare prediction, whereas the 2021 usage denotes a stacked-LSTM ensemble for index forecasting.