Solar flares are explosions in the solar atmosphere that release intense bursts of short-wavelength radiation and are capable of producing severe space-weather consequences. Flares release free energy built up in coronal fields, which are rooted in active regions (ARs) on the photosphere, via magnetic reconnection. The exact processes that lead to reconnection are not fully known and therefore reliable forecasting of flares is challenging. Recently, photospheric magnetic-field data has been extensively analysed using machine learning (ML) and these studies suggest that flare-forecasting accuracy does not strongly depend on how long in advance flares are predicted (Bobra & Couvidat 2015; Raboonik et al. 2017; Huang et al. 2018). Here, we use ML to understand the evolution of AR magnetic fields before and after flares. We explicitly train convolutional neural networks (CNNs) to classify SDO/HMI line-of-sight magnetograms into ARs producing at least one M- or X-class flare or as nonflaring. We find that flaring ARs remain in flare-productive states -- marked by recall >60% with a peak of ~ 80% -- days before and after flares. We use occlusion maps and statistical analysis to show that the CNN pays attention to regions between the opposite polarities from ARs and the CNN output is dominantly decided by the total unsigned line-of-sight flux of ARs. Using synthetic bipole magnetograms, we find spurious dependencies of the CNN output on magnetogram dimensions for a given bipole size. Our results suggest that it is important to use CNN designs that eliminate such artifacts in CNN applications for processing magnetograms and, in general, solar image data.

Solar flares - bursts of high-energy radiation responsible for severe space-weather effects - are a consequence of the occasional destabilization of magnetic fields rooted in active regions (ARs). The complexity of AR evolution is a barrier to a comprehensive understanding of flaring processes and accurate prediction. Though machine learning (ML) has been used to improve flare predictions, the potential for revealing precursors and associated physics has been underexploited. Here, we train ML algorithms to classify between vector-magnetic-field observations from flaring ARs, producing at least one M-/X-class flare, and non-flaring ARs. Analysis of magnetic-field observations accurately classified by the machine presents statistical evidence for (1) ARs persisting in flare-productive states --- characterized by AR area --- for days, before and after M- and X-class flare events, (2) systematic pre-flare build-up of free energy in the form of electric currents, suggesting that associated subsurface magnetic field is twisted, (3) intensification of Maxwell stresses in the corona above newly emerging ARs, days before first flares. These results provide new insights into flare physics and improving flare forecasting.