A Swin Transformer for Multiparameter Solar Wind Forecasting from MHD Simulations

Abstract Accurate forecasting of solar wind parameters is crucial for space weather forecasting and the protection of infrastructure. Existing physics-based magnetohydrodynamic (MHD) models are computationally expensive and often rely on assumptions. Most deep learning methods focus only on single-parameter predictions at spatial locations, making it difficult to characterize the multiparameter coupled dynamics that govern space weather effects. The model proposed in this study is based on the Transformer and adopts a Swin Transformer encoder–decoder architecture. It is trained on MHD simulation data and can predict eight solar wind parameters (number density, three velocity components, temperature, and three magnetic field components) on a two-dimensional polar coordinate grid of the solar equatorial plane. In order to handle the nonuniform spatial distribution of the grid in the polar coordinate system, this study introduces a position-dependent relative bias to encode the variation of physical spacing with heliocentric distance. The model was trained on MHD simulation data covering 631 Carrington rotations. In single-step (24 hr) predictions on the test set, the weighted correlation coefficients (wCC) exceed 0.98 for all parameters. In the autoregressive rolling prediction mode, the model maintains high accuracy for approximately 72 hr. When compared with OMNI observations covering the 2009–2023 testing period, the correlation coefficient for near-Earth radial velocity is 0.71. These results demonstrate that the Transformer-based architecture can effectively learn the multiparameter coupled dynamics of the solar wind from MHD simulations, serving as a computationally efficient alternative model and possessing the potential for operational evaluation.