Operational prediction of solar flares using a transformer-based framework.

Area of Science:

• Solar Physics
• Space Weather
• Machine Learning

Background:

• Solar flares are sudden energy releases from active regions (ARs) on the Sun.
• These events, along with coronal mass ejections, cause space weather, disrupting technologies like radio communication and power grids.
• Accurate solar flare prediction is vital for disaster risk management and technological preparedness.

Purpose of the Study:

• To develop and present SolarFlareNet, a transformer-based framework for predicting solar flares.
• To forecast the likelihood of active regions producing M5.0, M, or C-class flares within 24 to 72 hours.
• To enhance space weather forecasting capabilities through advanced machine learning.

Main Methods:

• Utilized a transformer-based deep learning architecture (SolarFlareNet) to model solar active region data as time series.
• Trained three separate transformers, each dedicated to predicting a specific solar flare class (M5.0, M, C).
• Incorporated magnetic parameters from Space-weather HMI Active Region Patches (SHARP) and flare data from NCEI flare catalogs (May 2010 - Dec 2022).

Main Results:

• Developed a fully operational SolarFlareNet system capable of near real-time solar flare predictions.
• Successfully modeled temporal dynamics in solar active region data using transformers for improved prediction accuracy.
• Extended the predictive model to a calibration-based probabilistic forecasting method.

Conclusions:

• SolarFlareNet provides accurate, early predictions of significant solar flares, crucial for mitigating space weather impacts.
• The transformer-based approach effectively captures complex temporal patterns in solar activity data.
• The system's operational status and web accessibility facilitate near real-time space weather monitoring and forecasting.