Improving solar flare forecasts with deep learning

The Sun, powered by nuclear fusion, convective motions, and dynamo processes, drives a range of dynamic phenomena that influence the entire heliosphere, collectively defining space weather. Among these, solar flares rank as some of the most powerful space weather events, characterized by explosive releases of magnetic energy emitting radiation across the electromagnetic spectrum and are often accompanied by solar energetic particles (SEPs). Such intense emissions pose risks to Earth’s technological infrastructure, including satellite communications, navigation systems, and power grids, as well as to human health in the context of space missions and high-altitude flights. Accurate and reliable solar flare prediction is therefore critical for effective space weather preparedness and risk mitigation. However, the complex dynamics underlying solar activity make precise forecasting a persistent challenge. This thesis investigates the application of deep learning — a branch of artificial intelligence (AI) that includes artificial neural networks (ANNs) and computer vision techniques — to advance solar flare prediction capabilities. Key contributions include (i) identification of limitations in existing performance evaluation methods and proposal of tailored evaluation methods to better address major challenges in flare forecasting models, (ii) introduction of a statistically robust cross-validation method for machine learning models trained on solar full-disk images, (iii) development of weaklysupervised learning frameworks to resolve spatial mislabelling issues, (iv) implementation of multimodal, spatio-temporal models leveraging coronal extreme ultraviolet (EUV) observations for enhanced predictive performances, and (v) prototyping of a generative AI model for probabilistic simulations of the solar corona evolution.