Benchmark study reveals a 15-year multimodal solar flare prediction dataset across 3,064 active regions, highlighting magnetic threshold impacts on forecast skill.
Key Points
To construct an extensive, coregistered 15-year multimodal dataset spanning 2011 through 2025 for advancing machine learning-based solar flare forecasting.
Coregistered 1,991,247 line-of-sight magnetogram crops across 3,064 solar active regions from SDO/HMI data, NOAA Solar Region Summaries, and NOAA X-ray flare records.
Extracted 29 magnetic features across FITS, PNG, CSV, and MP4 modalities with an 8:1:1 region-level split to eliminate temporal data leakage across partitions.
Evaluated predictive skill across seven forecasting windows and four GOES flare intensity thresholds using a Transformer and four baseline architectures.
Twenty-four-hour predictive models for ≥C1.0 and ≥M1.0 solar flare events demonstrated modest overall skill across all evaluated architectures.
A magnetic field saturation threshold of 1000 G served as an effective general baseline, though optimal saturation levels depended on the specific model and prediction task.