Solar Activity AI Forecaster

• Solar Activity AI Forecaster is a system that integrates real-time solar observations with hybrid physics-AI models to predict solar events across multiple time scales.
• It employs a dual data-model pipeline combining situational perception, in-depth analysis, and flare prediction modules to process multi-modal data efficiently.
• The forecaster has been rigorously verified using metrics like TSS and ROCA, demonstrating reliable near-real-time performance in operational settings.

The expression “Solar Activity AI Forecaster” has been used for a scalable dual data-model driven framework built on foundational models, integrating expert knowledge to autonomously replicate human forecasting tasks with quantifiable outputs in the Observation–Orientation–Decision–Action paradigm (Wang et al., 9 Aug 2025). In the broader literature, related AI systems forecast solar flares, active-region emergence, solar-wind speed, total and spectral solar irradiance, space weather indices, CME travel time, geomagnetic storm onset and recovery, and solar-cycle evolution, using SHARP magnetic parameters, full-disk magnetograms, EUV images, acoustic power maps, in situ solar-wind measurements, and sunspot records (Abduallah et al., 2024). An early operational milestone was an unmanned flare forecast service operating since August, 2015, providing 24-hour-ahead forecasts every 12 minutes, with realtime forecasting framed as the setting in which performance can be measured while confidently avoiding overlearning (Hada-Muranushi et al., 2016).

1. Conceptualization and historical development

A Solar Activity AI Forecaster is best understood as an operational or quasi-operational forecasting system in which machine learning or hybrid physics–AI models map solar or heliospheric observations to quantified predictions of future activity. In the 2025 formulation, the system is organized into three modules: a Situational Perception Module that generates daily solar situation awareness maps by integrating multi-modal observations, In-Depth Analysis Tools that characterize key solar features, and a Flare Prediction Module that forecasts strong flares for the full solar disk and active regions (Wang et al., 9 Aug 2025). The same paper describes a data flow of SPNet $\rightarrow$ IATools $\rightarrow$ FPNet, with total pipeline time $\le 6$ minutes on $1\times$ Tesla V100S GPU.

Operationalization predates that formulation. “DeepSun” presents a machine-learning-as-a-service framework for predicting solar flares on the Web, with a JSON API that includes POST /api/predict and POST /api/train, and reports typical latency $\sim 50$–$200$ ms per request (Abduallah et al., 2020). “A Deep Learning Approach to Operational Flare Forecasting” describes SolarFlareNet as fully operational and capable of making near real-time predictions of solar flares on the Web (Abduallah et al., 2024). These systems share an emphasis on regular data ingestion, explicit forecast horizons, and published verification metrics rather than offline classification alone.

This progression suggests that the defining feature of the Solar Activity AI Forecaster is not a single architecture, but the combination of near-real-time data flow, machine prediction, and operational verification.

2. Observational inputs and forecast targets

The forecast targets span multiple temporal and physical regimes. Flare-prediction systems use SHARP magnetic parameters, SDO/HMI magnetogram videos, AIA 304 $\unicode{x212B}$ images, GOES flare catalogs, and F10.7 radio flux (Abduallah et al., 2024). Active-region emergence models use Doppler shift and continuum intensity observations from HMI/SDO, convert 8-hour Dopplergram segments into acoustic power maps in four bands, and use a $9\times 9$ tiling scheme that yields 63 active tiles and a $252$-dimensional feature vector per time step (Kasapis et al., 2024). The 2025 superstorm study uses magnetogram cut-outs for morphology classification, HMI magnetogram videos for flare occurrence, LASCO cone-model parameters and solar wind for CME travel time, and L1 measurements for geomagnetic-storm alerts (Guastavino et al., 24 Jan 2025).

Medium-range forecasting extends the target set beyond flares. A multivariate ensemble model forecasts F10.7 cm, F15 cm, F30 cm, Ap, and Kp simultaneously over 27 days, using 108 days of prior indices and, when available, SDO/AIA image embeddings compressed by a $\sigma$-VAE from $\rightarrow$0 to 256 dimensions (Benson et al., 2021). A CNN-based solar-wind forecaster pairs AIA 193 $\rightarrow$1 images with ACE measurements at L1 through ballistic back-mapping (Raju et al., 2021). A Layer-Recurrent Network predicts total and spectral solar irradiance from HMI/SDO quicklook magnetograms and continuum images through 60 ring-wise filling factors derived from bipolar magnetic structures, penumbrae, and umbrae (Vieira et al., 2011).

Longer-horizon systems use historical sunspot records and dynamo surrogates. A WaveNet $\rightarrow$2 LSTM model forecasts 11 years of monthly averaged sunspot number and total sunspot area from historical time series (Benson et al., 2020). Reservoir computing has been applied to the 13-month-smoothed International Sunspot Number (Espuña-Fontcuberta et al., 2020). A diffusion-dominated flux-transport dynamo has been coupled to a NARX network for polar-field reconstruction and cycle forecasting (Macario-Rojas et al., 2018), while a nonlinear mean-field $\rightarrow$3 dynamo with NEMPI has been combined with a two-layer NARX correction scheme for 1-, 6-, 12-, and 18-month forecasts (Kleeorin et al., 2024). A more recent intermediate-timescale system models “space weather seasons” of approximately 6 to 18 months with hemispheric SARIMA baselines, asymmetric Gaussian burst injection, and random-forest regression for amplitude and duration (Shetye et al., 26 May 2026).

Taken together, these studies define the Solar Activity AI Forecaster as a family of multi-horizon systems rather than a flare-only forecaster.

3. Model families and mathematical structure

The literature includes classical tabular classifiers, recurrent sequence models, convolutional image models, transformers, masked autoencoders, and physics-guided hybrids. In flare prediction from magnetic parameters, support vector machines, multilayer perceptrons, and random forests remain important baselines and, in some studies, leading methods. One SVM study uses 25 scalar parameters from HMI vector magnetic field data and finds that the top 4 features already yield nearly the best TSS, while the optimal subset for TSS is the top 13 features (Bobra et al., 2014). Another study computes 13 SHARP-derived predictors, including Schrijver’s $\rightarrow$4, FSPI, TLMPIL, DI, WL$\rightarrow$5, and Ising-energy variants, and reports that random forests could be the prediction technique of choice for that sample (Florios et al., 2018).

Sequence models dominate multistep forecasts. The multivariate space-weather index system uses a stacked LSTM with three recurrent layers and 512 hidden units per layer, mapping a 108-day sequence to a direct $\rightarrow$6-day multivariate forecast $\rightarrow$7 (Benson et al., 2021). Its ensemble uncertainty is defined by

$\rightarrow$8

The active-region-emergence model uses 3 stacked LSTM layers, each with 64 hidden units, on 80-hour acoustic-power histories to predict whether continuum intensity values will decrease 5 hours ahead (Kasapis et al., 2024).

Image-driven systems employ CNNs or transformers. The solar-wind model uses a custom CNN with three convolutional blocks and two dense layers on downsampled $\rightarrow$9 AIA 193 $\le 6$0 images (Raju et al., 2021). SolarFlareNet combines Conv1D, a single-layer LSTM, and $\le 6$1 Transformer Encoder Blocks over a sequence length $\le 6$2 at 12-minute cadence, with $\le 6$3 selected SHARP parameters (Abduallah et al., 2024). The 2025 “Solar Activity AI Forecaster” uses a multi-modal Masked Autoencoder plus transformer encoder in SPNet, and an Efficient Masked Autoencoder for Flare Prediction with Physical Prior-guided Adaptive Masking in FPNet (Wang et al., 9 Aug 2025). The May 2024 superstorm study uses a Vision Transformer for active-region morphology and an LRCN for 24-hour magnetogram videos (Guastavino et al., 24 Jan 2025).

Hybrid physics–AI systems are equally prominent. Reservoir computing is used as a model-free, data-based forecaster for sunspot cycle 25 (Espuña-Fontcuberta et al., 2020). Mean-field dynamo and NARX combinations appear in both long- and short-range cycle forecasts (Macario-Rojas et al., 2018). The 2024 short-range cycle predictor defines a NARX input vector $\le 6$4 from four past observations and four model outputs, then computes

$\le 6$5

with 24 hidden neurons (Kleeorin et al., 2024). The May 2024 CME-arrival model uses a physics-driven neural cascade in which $\le 6$6 estimates the drag parameter and $\le 6$7 predicts travel time under a joint MSE-plus-physics loss (Guastavino et al., 24 Jan 2025).

A common structural pattern is therefore the coupling of data-rich neural representations to physically constrained variables, priors, or post-processing.

4. Operational workflows and service architecture

Operational workflows are defined by cadence, automation, and reproducibility. The 2016 realtime flare forecast service provides 24-hour-ahead forecasts every 12 minutes and emphasizes unmanned operation (Hada-Muranushi et al., 2016). SolarFlareNet uses a daily batch job to retrieve the most recent SHARP parameters for all visible active regions through a SunPy/JSOC API, forms a 10-sample window per region, and delivers the daily forecast for approximately 100 active regions in less than 1 minute after data retrieval, with predictions refreshed once per 24 hours (Abduallah et al., 2024). DeepSun exposes prediction and training services through a REST-capable interface and supports file upload, pretrained or custom models, and remote programming users (Abduallah et al., 2020).

The dual data-model framework in 2025 extends this operational logic to full-disk multimodal perception, entity extraction, and flare prediction in one pipeline (Wang et al., 9 Aug 2025). Its preprocessing includes FITS-to-JPG conversion, disk-center alignment, Hough-transform circle detection for H$\le 6$8 images, intensity normalization, and multi-modal alignment onto a shared $\le 6$9 grid. IATools then performs AR extraction with DBSCAN, CH/FL connectivity labeling, and physical-parameter inference including Carrington longitude, $1\times$0-value, Hale type, McIntosh classification, and Flare Index.

Other systems define operational pipelines for non-flare tasks. The 27-day multivariate forecaster uses a direct-forecast strategy that outputs all 27 days at once rather than rolling forward step-by-step (Benson et al., 2021). The solar-wind system uses ballistic back-mapping to pair each 2-hour-cadence AIA image with an ACE speed at the corresponding arrival time (Raju et al., 2021). The irradiance model ingests HMI/SDO quicklook products with latency $1\times$1 minutes and emits irradiance forecasts up to three days in advance (Vieira et al., 2011).

These examples show that service architecture is a central part of the subject: the forecaster is not merely a trained model, but a continuously executing data-and-decision pipeline.

5. Verification metrics and reported results

Forecast verification in this literature is dominated by contingency-table metrics, regression errors, and probabilistic scores. A recurrent formula is the True Skill Statistic,

$1\times$2

which is used in flare and geomagnetic-storm studies (Guastavino et al., 24 Jan 2025). Probabilistic systems additionally report Brier Score, Brier Skill Score, ROCA, and calibration diagnostics (Abduallah et al., 2024).

Forecast task	System	Reported result
Full-disk and AR flare forecasting	“Solar Activity AI Forecaster”	Full-disk: $1\times$3, $1\times$4, $1\times$5; AR-level: $1\times$6, $1\times$7, $1\times$8 (Wang et al., 9 Aug 2025)
Operational flare forecasting	SolarFlareNet	24 h deterministic $1\times$9: $\sim 50$0 ($\sim 50$1), $\sim 50$2 ($\sim 50$3), $\sim 50$4 ($\sim 50$5); operational accuracy: $\sim 50$6 (24 h), $\sim 50$7 (48 h), $\sim 50$8 (72 h) (Abduallah et al., 2024)
$\sim 50$9 and $200$0 flare prediction	Random forests on 13 predictors	$200$1: $200$2, $200$3, $200$4 at probability threshold $200$5; $200$6: $200$7, $200$8, $200$9 at probability threshold $\unicode{x212B}$0 (Florios et al., 2018)
27-day multivariate space-weather indices	LSTM ensemble with image fusion	$\unicode{x212B}$1–$\unicode{x212B}$2 improvement of the root-mean-square error while including solar image data with time series data compared to using time series data alone; stacked LSTM reduces RMSE by $\unicode{x212B}$3–$\unicode{x212B}$4 relative to persistence/running-average across all five indices (Benson et al., 2021)
Solar-wind speed forecasting	CNN-based deep learning	$\unicode{x212B}$5 km/s, $\unicode{x212B}$6, threat score $\unicode{x212B}$7 with zero false alarms (Raju et al., 2021)
May 2024 CME and storm forecasting	ViT + LRCN + physics-driven cascade + LSTM	$\unicode{x212B}$8 h vs. observed $\unicode{x212B}$9 h, $9\times 9$0 h; storm-onset forecast: $9\times 9$1 detection rate, false-alarm ratio $9\times 9$2 (Guastavino et al., 24 Jan 2025)

Additional domains report similarly specific results. The short-term irradiance forecaster obtains 24-hour forecast error lower than $9\times 9$3 in the band from 115 to 180 nm, while the model error can reach $9\times 9$4 in the band from 180 to 310 nm (Vieira et al., 2011). The WaveNet $9\times 9$5 LSTM cycle predictor reports validation RMSE $9\times 9$6 for sunspot number and $9\times 9$7 for total sunspot area (Benson et al., 2020). The dynamo $9\times 9$8 NARX short-range cycle forecaster reports RMS$9\times 9$9, RMS$252$0, RMS$252$1, and RMS$252$2, lower than the listed McNish–Lincoln, Standard Method, and Combined Method baselines (Kleeorin et al., 2024).

The verification literature therefore indicates substantial task dependence: high flare-skill scores do not directly translate to solar-wind, irradiance, or solar-cycle prediction, and cross-task comparison requires care.

6. Limits, misconceptions, and research directions

A common misconception is that offline accuracy is sufficient. The operational flare-forecast paper from 2016 states that only by realtime forecast can performance be experimentally measured while confidently avoiding overlearning (Hada-Muranushi et al., 2016). Another misconception is that more modalities uniformly improve every target. In the multivariate index study, solar-proxy $252$3, $252$4, and $252$5 RMSE drops by an additional $252$6–$252$7 when solar-image features are included, whereas geomagnetic indices see comparable or slightly degraded performance (Benson et al., 2021).

A further misconception is that AI can dispense with physics. Several papers argue the opposite by construction. The solar-wind CNN relies on ballistic back-mapping, but explicitly notes that the constant-$252$8 approximation ignores acceleration, stream–stream interactions, and Parker-spiral geometry (Raju et al., 2021). The CME-arrival model uses a physics-driven cascade rather than a purely black-box regressor, yet the same study also states that its superstorm analysis is a single case-study and needs multi-event validation (Guastavino et al., 24 Jan 2025). The 2025 modular forecaster identifies residual segmentation or parameterization errors as a source of error propagation and proposes uncertainty propagation and Bayesian calibration as future work (Wang et al., 9 Aug 2025).

Long-horizon prediction is especially contested. In “Forecasting the solar activity cycle: new insights,” significant turbulent pumping severely degrades the memory of the dynamo, so long term prediction of the solar cycle is not possible; only a short term prediction of the next cycle peak may be possible based on observational data assimilation at the previous cycle minimum (Nandy et al., 2013). This caution is consistent with other cycle-forecast studies that note error accumulation when predictions are fed back as inputs (Benson et al., 2020).

The published research directions are correspondingly hybrid. Proposed extensions include Bayesian neural networks, multi-model ensembles, uncertainty quantification by larger ensembles, additional exogenous indices such as polar field strength and F10.7 cm radio flux, data assimilation, and expansion from flare forecasting to CME and SPE forecasting modules (Wang et al., 9 Aug 2025). This suggests that the most durable form of the Solar Activity AI Forecaster is likely to remain a modular system in which expert knowledge, physical priors, and machine learning are jointly updated rather than cleanly separated.