Speaker
Description
Early identification of flare-productive solar active regions is essential for operational space-weather forecasting. We present an integrated deep-learning framework for automated active-region detection, localization, magnetic classification, and short-term flare forecasting, designed for continuous monitoring and near-real-time deployment.
The system combines three complementary components. SUN-FD operates directly on full-disk magnetograms, performing simultaneous active-region detection, localization, and magnetic classification via deep learning–based object detection. It outputs bounding boxes together with probabilistic class scores, enabling automated full-disk surveillance with uncertainty-aware estimates.
SUN-ARC refines magnetic characterization by classifying extracted active-region patches into Mount Wilson classes using supervised deep learning. The model captures polarity configurations and morphological complexity and provides calibrated probabilistic outputs suitable for downstream decision-making.
Building on these products, DFF (DeepFlareForecast) predicts the likelihood of flares within the next 24 hours from multi-wavelength, time-ordered sequences of active-region observations (SDO/HMI magnetograms jointly with SDO/AIA channels). The forecasting model follows a spatio-spectro-temporal design: a shared convolutional encoder extracts compact spatial representations per timestep, channel-aware fusion emphasizes the most informative wavelengths, and a lightweight temporal transformer captures the evolution of magnetic and coronal signatures leading to flares. The model produces probabilistic forecasts to reflect flare hierarchy and support threshold-based alerting.
Together, these tools provide a coherent operational pipeline from full-disk magnetic monitoring to active-region characterization and flare alert generation, supporting timely mitigation strategies for space-based and ground-based technological systems.
| Numerical model | SUN-FD, SUN-ARC, DFF |
|---|