S$^3$N: A Spherical Spiral Scanning Network for Weather Forecasting

Fan Yan, Chen Hui, Weisi Lin, Haiqi Zhu, Feng Jiang, Sun-Yuan Kung, Wei Zhang

arXiv preprint · 03 October 2026 · Nanjing University of Information Science & Technology (NUIST), Princeton University

In one sentence. S3N uses an L2 projection between latitude-longitude and HEALPix grids and spiral state-space scanning to forecast weather at 4, 7 and 10 days with slower error growth.

Abstract

Machine learning-based weather prediction (MLWP) has achieved strong performance in global weather forecasting. Recent Hierarchical Equal Area isoLatitude Pixelation (HEALPix)-based methods use the HEALPix (HP) grid to avoid area distortion near the poles of conventional latitude-longitude (LL) grids. However, existing HP-based approaches often use pointwise mapping methods and process HP pixels within separate base faces or local windows. Consequently, the mapping may introduce reconstruction errors and cross-face communication depends on handcrafted boundary handling or shifted windows. We propose the Spherical Spiral Scanning Network (S$^3$N) to address both limitations. First, L2Proj provides a bidirectional method for mapping atmospheric fields between the LL and HP grids through an $L^2$ projection of their continuous finite-element representations. Second, the Attention-Guided Quad-Spiral State-Space Scanning (AQSS) block uses cross-latitude attention to guide selective state-space updates along four global pole-to-pole spiral paths. This design enables continuous information propagation across HP base-face boundaries without additional boundary-processing mechanisms. Experiments show that S$^3$N achieves better results at 4-, 7-, and 10-day lead times, and exhibits slower error growth in long-range forecasting.

Labels

Labels and summary are generated by Claude and reviewed weekly. Wrong? Edit this entry.