Deep-Learning-Based Models for Solar Flare Forecasting

One of the most important discoveries of the past century in solar physics was the detection of magnetic fields in sunspots by G. E. Hale, followed by the establishment of the existence of the much weaker global magnetic field of the Sun. Since then, the structure and polarity of magnetic fields in magnetically active regions (ARs) have been extensively investigated. Among the most prominent manifestations of solar magnetic activity are solar flares. Solar flares are localized explosive phenomena observed in the solar atmosphere. First observed in white light by R. Carrington, they are characterized by enhanced radiation across the electromagnetic spectrum. Flares are thought to originate from the rapid conversion of magnetic energy stored in stressed coronal fields and may be associated with other manifestations of solar activity, including solar energetic particle (SEP) events and, in some cases, coronal mass ejections (CMEs). Because flare electromagnetic radiation reaches the near-Earth environment within minutes, major flares can produce prompt ionospheric disturbances and affect radio communication, satellite navigation, and other technological systems that depend on the stability of the geospace environment. Despite their operational relevance, reliable flare forecasting remains challenging. Flare productivity depends on the nonlinear evolution of magnetic structures across multiple spatial and temporal scales, while the observational data are high-dimensional, heterogeneous, and strongly affected by the rarity of major events. The Solar Dynamics Observatory (SDO) era has provided continuous, high-cadence observations of the solar atmosphere at multiple wavelengths, together with photospheric magnetic-field and Doppler measurements, and physical parameters derived from vector magnetic-field data. At the same time, recent developments in deep learning have made it possible to analyze such large observational datasets and extract relevant information. These developments motivate the central objective of this thesis: the investigation of deep-learning-based flare forecasting methods. Since solar flares are understood to be magnetically driven phenomena, active-region magnetograms constitute a natural starting point and a widely used source of predictive information. While many forecasting approaches rely on single-time representations of active-region magnetic field, this thesis focuses on the forecasting value of their temporal evolution and investigates how predictive performance is affected by progressively incorporating further sources of information, including continuum and extreme-ultraviolet (EUV) observations, as well as physical parameters derived from vector magnetic-field data. The work is developed through four connected phases. First, convolutional autoencoders are used to learn compact representations of line-of-sight (LoS) magnetograms of ARs, with the aim of assessing whether the latent space preserves physically relevant information about AR magnetic morphology. Second, sequences of encoded LoS magnetograms are organized as temporal sequences and used as inputs to recurrent classifiers to predict whether an AR will produce an M- or X-class flare within the following hours, based on its prior evolution, thereby assessing the predictive power of both the learned representations and their temporal evolution. Third, the forecasting framework is extended to a multimodal and multiwavelength setting that combines LoS magnetograms, continuum and EUV observations, and scalar physical parameters derived from vector magnetograms. In this case, the classification target is redefined as flares of class C5.0 or above, and the temporal analysis is restricted to 36 min of observations with a 2 h prediction window. Here, the focus shifts from the long-term evolution to short-term forecasting, where rapidly evolving photospheric and coronal conditions may be interpreted as preflare signatures. Finally, Layer-wise Relevance Propagation (LRP) is applied to interpret the predictions of this multimodal model and to examine how relevance is distributed across input modalities, as well as whether relevance allocation is associated with physically meaningful structures, such as polarity inversion lines (PILs) or regions of enhanced magnetic non-potentiality. A first key innovation of this thesis is the use of compact neural representations learned through magnetogram reconstruction as inputs for flare forecasting. The results show that these representations can preserve both large-scale magnetic organization and smaller-scale magnetic structures, and that, when organized as temporal sequences, they also retain information on the temporal evolution of physical parameters derived from LoS magnetograms. Reconstruction objectives that explicitly constrain local magnetic-field gradients improve the recovery of gradient-sensitive physical diagnostics, indicating that the design of the learning objective affects the physical consistency of the latent representation. When LoS magnetogram representations are arranged as time series and used for forecasting, temporal modeling provides strong discrimination between flare-productive and non-flaring sequences, supporting the idea that flare-relevant information is encoded not only in the instantaneous magnetic configuration, but also in its evolution. A second key innovation is the extension of the forecasting framework to a multimodal and multiwavelength setting. The multimodal forecasting experiments show that predictive performance improves as additional sources of observational and physical information are incorporated. Models based only on LoS magnetograms already capture useful predictive information, but the inclusion of continuum and EUV imagery provides complementary information about the photospheric and upper-solar-atmosphere evolution of ARs. The best performance is obtained when image-based representations are combined with physical parameters describing global magnetic properties. The interpretability analysis indicates that the model decisions are often supported by localized relevance patterns in photospheric, transition-region, and coronal observations, with qualitative correspondence to physically meaningful photospheric magnetic structures, such as PILs, strong magnetic-field gradients, enhanced current helicity, and regions of increased free magnetic energy density, as well as to upper-solar-atmosphere structures traced by EUV emission, including localized brightenings and coronal loops. The relevance-allocation analysis further shows that image relevance generally accounts for most of the total relevance budget, while physical-parameter relevance varies across classification outcomes and reflects different contributions from magnetic flux and current helicity. Misclassified cases reveal complementary failure modes: false positives are often linked to visually complex structures that resemble flare-productive morphology, while false negatives may correspond to regions in which physically relevant signatures are present but too localized or insufficiently emphasized by the learned image representation. These findings show that attribution methods are useful not only for interpreting successful predictions, but also for identifying the limitations and failure modes of deep-learning-based flare forecasting models. This thesis contributes to solar flare forecasting by developing a unified deep-learning framework from data representation learning to interpretability. The results support the central conclusion that solar flare forecasting can benefit from frameworks able to integrate spatial structure, temporal evolution, multiwavelength observations, and physically motivated magnetic-field diagnostics. At the same time, the analysis highlights the importance of evaluating forecasting models not only in terms of predictive skill, but also in terms of calibration, robustness, physical consistency, and interpretability.