Speaker
Description
Operational solar flare forecasting requires computationally efficient and energy-optimal methods that maximize the use of available observational resources to deliver timely and reliable predictions. Synoptic full-disk observations from the Solar Dynamics Observatory (SDO) provide continuous monitoring of solar magnetic activity over more than one solar cycle, enabling detailed studies of solar variability and space weather impacts. The Space-weather HMI Active Region Patches (SHARP) vector magnetic field (VMF) maps and parameters, derived from the Helioseismic and Magnetic Imager (HMI), support investigations of active region evolution and flare triggering mechanisms. In this study, we use time series of SHARP VMF maps as input to a Disentangled Variational Autoencoder (VAE), a Disentangled Representation Learning (DRL) method that extracts low-dimensional features capturing the morphological and dynamic characteristics of active regions. These VAE-derived features exhibit temporal evolution patterns similar to, but not redundant with, certain SHARP parameters, indicating that their combination provides an enhanced representation of solar magnetic activity. We construct a joint dataset merging human-curated SHARP parameters with machine-learned VAE features, resulting in a high-fidelity input for flare forecasting. Our forecasting pipeline utilizes this dataset to produce binary (Flare vs. No-Flare, Alert vs. No-Alert) and multi-class probabilistic predictions. The pipeline employs a Long Short-Term Memory (LSTM) network to learn the temporal evolution of the features for several time windows, followed by logistic regression to estimate probabilities for strategically labeled event classes. This integrated approach highlights the value of combining physics-derived and machine-learned representations to improve the accuracy and robustness of solar flare forecasting models.
| Do you plan to attend in-person or online? | In-person |
|---|
