StretchCast: Global-Regional AI Weather Forecasting on Stretched Cubed-Sphere Mesh

Abstract

Global AI weather forecasting still relies mainly on uniform-resolution models, making it hard to combine regional refinement, two-way regional-global coupling, and affordable training cost. We introduce StretchCast, a global-regional AI forecasting framework built on a variable-resolution stretched cubed-sphere (SCS) mesh that preserves a closed global domain while concentrating resolution over a target region. Within this framework, we develop a one-step predictor, SCS_Base Model, and a rollout-oriented multistep predictor, SCS_FCST4 Model, to test the feasibility of SCS-based forecasting and the benefit of joint multistep training. Experiments use ERA5 with 69 variables over 1998-2022. Because training compute remains limited, this study uses a coarse-resolution proof-of-concept configuration rather than a final high-resolution system. Even with only about 7,776 effective global grid cells and roughly 0.875 degree resolution over the center-refined face, the 23M-parameter SCS_Base Model yields stable multivariate forecasts. With 83M parameters and training cost on the order of hours, SCS_FCST4 Model delivers competitive medium-range anomaly-correlation evolution over the target region after unified reprojection, especially for geopotential height, specific humidity, and part of the lower-tropospheric winds, while maintaining smooth cross-face continuity and realistic multiscale structure in typhoon and spectral analyses. These results support StretchCast as a practical lightweight foundation for global-regional AI weather forecasting.

Figures (21)

• Figure 1: Schematic of the stretched cubed-sphere (SCS) configuration used in this study and its core-halo interaction mechanism, with the center-refined face placed over eastern China. (a) Global view of the SCS grid and the center-refined region; (b) three-dimensional view of the sphere partitioned into six faces; (c) cross-shaped flattened arrangement of the six faces; and (d) core-halo structure of a single face, where the blue block denotes the face core, the surrounding colored bands denote halos, and N and N+2h indicate the corresponding numbers of grid points along the edges. Black arrows indicate the mapping from the spherical grid to the faceted representation and then to the flattened tensor representation.
• Figure 2: Comparison between an idealized latitude-longitude field and its reconstruction from the SCS representation. (a) Original latitude-longitude field; (b) field reconstructed back from the SCS representation, with red curves marking face boundaries and red numbered boxes indicating face indices; and (c) absolute reconstruction error, with black curves marking face boundaries, a color bar showing the error magnitude, and the maximum absolute error reported in the panel title.
• Figure 3: Architecture of StretchCast. (a) Pipeline of the SCS_Base one-step model. Initial Fields @ T (SCS Mesh) are processed by an Encoder, a Face Mixer, a Global Mixer, and a Decoder to produce Forecast Fields @ T+1 (SCS Mesh). Pink modules denote the Core-Halo Interaction System, and blue modules denote 3D Spherical Position FiLM. (b) Pipeline of the SCS_FCST4 multi-step model. Initial Fields @ T-1 and Initial Fields @ T are encoded by a shared Encoder, processed by a Temporal Mixer and the same spatial backbone, and then passed to Split to 4 steps and multiple Decoders to output Forecast Fields @ T+1, T+2, T+3, and T+4.
• Figure 4: Denormalized spatial RMSE distribution of the SCS_Base one-step model over the test set at +6 h. Panels (a-e) show Z200, Q200, T200, U200, and V200; panels (f-j) show Z500, Q500, T500, U500, and V500; panels (k-o) show Z850, Q850, T850, U850, and V850; panel (p) shows SLP; panel (q) shows T2m; panel (r) shows U10m; and panel (s) shows V10m. The color bar below each panel gives the RMSE magnitude in the corresponding physical unit.
• Figure 5: Forecast of 850 hPa specific humidity (Q850) for a representative case from the SCS_Base one-step model. Panels (a-e) show the ground truth (GT) at +6 h, +24 h, +48 h, +96 h, and +120 h; panels (f-j) show the prediction (Pred) at the same lead times; and panels (k-o) show the absolute error (Abs. Err.) at the same lead times. The first two rows share the same field color scale, the third row uses an absolute-error color scale, and the text in the lower-left corner reports the experiment name and sample index. If highlighted boxes are present, they indicate two additionally annotated local weather-process regions during this period: the high-moisture region over the Pearl River Basin and the atmospheric river (AR) event extending from the eastern Pacific to North America.