Prediction of Time-evolving Radial Magnetic Fields on the Solar Surface Using Deep Learning

Abstract We develop two artificial intelligence–based models (Models A and B) to predict time-evolving photospheric magnetic fields with an adjustable time step, ranging up to one solar rotation later. Model A predicts future magnetic field data using three consecutive radial magnetic field data sets, each 12 hr apart. Model B reconstructs evolving magnetic fields using two sets of three consecutive data sets, one from a given date and the other from 27 days later, each 12 hr apart. To train and evaluate our models, we use a Pix2PixCC-based architecture and radial magnetic field data sets from the Solar Dynamics Observatory/Helioseismic and Magnetic Imager during the solar maximum periods of 2012–2016 and 2021–2023. Models A and B successfully generate magnetic field data corresponding to the input prediction time step. Based on quantitative comparison, Model A outperforms the persistence model and performs comparably to the classical surface flux transport model. Model B shows improved performance compared to both models. We also compare predictions for National Oceanic and Atmospheric Administration Active Regions 12673 and 13664, which rapidly emerged and produced powerful solar flares. Notably, both Models A and B can predict increases in total unsigned magnetic fluxes several days in advance if their evolution is already captured in the input data. Model B also reconstructs radial magnetic fluxes that smoothly overlap between the input data and the model outputs. These results demonstrate the ability of our models to generate time-evolving magnetic field data flexibly and capture emerging magnetic fluxes, which are crucial for heliophysics research and space weather forecasting.