Logo

Authors

Jeong, Hyun-Jin

Issue Date

2025-11

Citation

논문 The Astrophysical Journal Supplement Series, v.281, no.1, pp.23-

Journal Title

The Astrophysical Journal Supplement Series

Volume

281

Number

1

Start Page

23

DOI

10.3847/1538-4365/ae13a2

Abstract

Accurately determining solar wind parameters is crucial for Sun–Earth space research, as they significantly affect spacecraft safety and ground-based power systems. Traditionally, solar wind conditions are derived using coupled coronal and heliospheric models, with the latter initialized by the former’s output at 0.1 au, a computationally intensive and time-consuming process that limits real-time space weather forecasting. In this work, we propose a machine-learning-based method for generating solar wind parameters at 0.1 au. Specifically, we employ a U-Net neural network, trained using the output of the COolfluid COroNal UnsTructured (COCONUT) model as the learning target and Global Oscillation Network Group–Air Force Data Assimilative Photospheric Flux Transport magnetograms as input. The model achieves correlation coefficients of 0.992 for radial velocity, 0.987 for number density, and 0.991 for radial magnetic field on the test set, with derived Alfvén speed and dynamic pressure reaching 0.996 and 0.769, respectively, demonstrating strong capability in reconstructing key solar wind parameters. Moreover, the model effectively captures the temporal evolution of these parameters within a single Carrington rotation. Once trained, the model generates full-surface solar wind predictions at 0.1 au in 7.8 s on a CPU-only device and 0.065 s on a cluster with one GPU and 10 CPU cores, achieving 15× and 1800× speed-ups, respectively, over the COCONUT magnetohydrodynamic simulation, which requires at least 1 hr to obtain a converged steady-state solution and over 2 minutes on 288 CPU cores per prediction.