Prediction of geomagnetic events from solar wind data using deep learning

The recent technological maturity attained by deep learning drew the attention of numerous scientific communities. Among these, Space Weather is leveraging such tools to support its activities related to forecasting harmful events. Coronal mass ejections (CMEs) are one of the most critical phenomena occurring in the solar system: their propagation may impact the Earth, altering the equilibrium of the terrestrial surface under different aspects, requiring their prediction to take counter- measures accordingly. Classical forecasting methods are built upon solar remote-sensing observations to forecast the CME onset, intensity, and arrival time. Although such methods could provide alerts within 1-4 days in advance, their estimations are affected by large uncertainties. On the other hand, deep learning has been offering valid alternatives through recent studies, devising data-driven models to obtain real-time alerts while monitoring such events remotely. The goal of this work is that of developing neural network architectures able to offer CME predictions leveraging the La- grangian point L1 measurements, taking advantage of historical data related to these phenomena past behaviors to predict future trends. In this paper, two main phases may be distinguished: first, the manipulation of the dataset - mostly through augmentation techniques - to make it more suitable for the proposed prediction steps. Second, the implementation of different network structures for multiple classification tasks concerning various aspects of CMEs, to prove the effectiveness of deep learning algorithms in reaching the desired goal.