Abstract Flare precursors can indicate where and when flares occur in active regions, yet their spatial and temporal characteristics remain poorly understood. The Mg ii 2798.82 Å triplet has been identified as a strong flare precursor, likely linked to temperature or density changes at various chromospheric heights. We aim to study the locations of flare precursors relative to later flare sites using Interface Region Imaging Spectrograph observations and machine learning and to quantify the significance of Mg ii triplet emission before flares. We developed a flare location algorithm combining a variational autoencoder to detect spectra dissimilar to quiet-Sun profiles and an XGBoost classifier to identify flaring sites based on UV brightness. Additionally, we trained a convolutional neural network (CNN) to detect precursor signatures during the preflare phase and tracked the evolution of the Mg ii triplet to capture potential precursor-specific shapes beyond single-peaked emission. The CNN predicted flare locations in 79% of flaring active region observations. In 71% of these, the Mg ii triplet was already in emission at the future flare site, independent of intensity changes in Mg ii h&k and Si iv 1403 Å. However, we found no clear temporal patterns or causal connections between precursors and flare onset. Our models robustly identify flare-prone locations, many showing Mg ii triplet emission hours in advance. Nonetheless, there is no consistent spectral evolution pinpointing when a flare will occur. This suggests that incorporating chromospheric magnetic field information may be key to distinguishing flare-related processes from unrelated small-scale reconnection events.
Zbinden et al. (Wed,) studied this question.