Sparse Residual Routing Boosts Efficiency and Accuracy in Weather Prediction

Modern Vision Transformer (ViT)-based weather forecasting models often apply uniform computation across all spatial tokens, even though many atmospheric grid points are redundant or evolve smoothly. This leads to inefficient processing. Traditional token-efficiency methods like pruning or merging are unsuitable for dense spatiotemporal prediction tasks like weather forecasting, where every grid cell needs a physically meaningful representation. To address this, the paper proposes Sparse-Reslim, a novel, parameter-free plug-in routing module. Sparse-Reslim intelligently routes only 25% of spatial tokens through the computationally intensive middle transformer blocks. Crucially, it treats these blocks as residual updates, computing only the change for the routed tokens and scattering this delta back to the full sequence. Unselected tokens retain their original representations, ensuring no data loss or grid cell modification. This approach yields substantial benefits: training speedups of up to 3.18x and over 2.2x lower peak memory usage at high resolutions. Remarkably, Sparse-Reslim also improves forecast accuracy across all evaluated variables and model families (deterministic Transformer and diffusion models), demonstrating that sparse processing can lead to both efficiency and enhanced performance in complex dense prediction tasks.