A Convolutional Neural Network–derived Catalog of Solar Flares from Soft X-Ray Observations

Abstract A convolutional neural network (CNN) is used to construct a new catalog for solar flares based on high-resolution (1 s cadence) Geostationary Operational Environmental Satellites (GOES) soft X-ray data. The CNN is trained to identify flare rise episodes. From 2018 January 1 to 2025 August 22, the algorithm detects 111,580 flare candidates, compared with 14,612 events in the corresponding GOES catalog. For each candidate, the probability of being a true positive is quantified by Bayesian inference based on the peak flux, rise time, and temporal coincidence with cataloged events where available. The flare size and waiting-time distributions are studied and compared with the GOES catalog. The CNN catalog shows a steeper power-law index for raw peak fluxes (−2.59 ± 0.02) than GOES (−2.25 ± 0.04), indicating the CNN’s higher sensitivity to small events. After background correction, the indices are −1.97 ± 0.02 (CNN) and −2.05 ± 0.04 (GOES). The CNN catalog extends the power-law distribution of flare peak fluxes by 1 order of magnitude at the small-flux end compared with the GOES background-subtracted catalog. A Bayesian block analysis of the waiting-time distributions from the GOES and CNN catalogs indicates broad consistency with a piecewise Poisson process. We find that previously reported correlations between flare sizes and waiting times are significantly influenced by obscuration, that is, undercounting weaker or overlapping flares during periods of elevated flux. The new CNN catalog provides a foundation for complete and consistent studies of solar flare statistics.