Swin Transformer-based encoder–decoder architecture. The model divides the input two-dimensional solar wind parameter field into patches and processes them through hierarchical encoder and decoder stages. Downsampling and upsampling modules connect different levels, with skip connections fusing encoder features and decoder outputs. Each stage consists of Swin Transformer blocks employing shifted window attention and position-dependent relative biases for polar coordinate grids. The polar coordinate plots at the four corners are schematic diagrams illustrating the polar coordinate grid geometry corresponding to the input field X^(t), the token representations H^(t) and H^(t+Δt), and the predicted field ﹩{\hat{{\boldsymbol{X}}}}^{(t+{\rm{\Delta }}t)}﹩. The plots attached to H^(t) and H^(t+Δt) do not represent the learned feature values; these features reside in an abstract embedding space and do not have a direct physical-space interpretation.