Title: Deep learning reconstruction of sunspot vector magnetic fields for forecasting solar storms

Abstract: Solar magnetic activity produces extreme solar flares and coronal mass ejections, which pose grave threats to electronic infrastructure and can significantly disrupt economic activity. It is therefore important to appreciate the triggers of explosive solar activity and develop reliable space-weather forecasting. Photospheric vector-magnetic-field data capture sunspot magnetic-field complexity and can therefore improve the quality of space-weather prediction. However, state-of-the-art vector-field observations are consistently only available from Solar Dynamics Observatory/Helioseismic and Magnetic Imager (SDO/HMI) since 2010, with most other current and past missions and observational facilities such as Global Oscillations Network Group (GONG) only recording line-of-sight (LOS) fields. Here, using an inception-based convolutional neural network, we reconstruct HMI sunspot vector-field features from LOS magnetograms of HMI as well as GONG with high fidelity (~ 90% correlation) and sustained flare-forecasting accuracy. We rebuild vector-field features during the 2003 Halloween storms, for which only LOS-field observations are available, and the CNN-estimated electric-current-helicity accurately captures the observed rotation of the associated sunspot prior to the extreme flares, showing a striking increase. Our study thus paves the way for reconstructing three solar cycles worth of vector-field data from past LOS measurements, which are of great utility in improving space-weather forecasting models and gaining new insights about solar activity.