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

Jin Feng

physics.ao-phMar 28, 2026

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

Authors: Jin Feng

Organizations: Institute of Urban Meteorology, China Meteorological Administration (IUM, CMA), Beijing, China · LeadSee Team, Beijing Meteorological Service

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 & tables

Organization	Model	Resolution	Num. grids	Meteo vars.	Params.	Computation resources for training	Time for training
Huawei	Pangu	0.25 (Global)	1110K	69	0.256B	Nvidia V100 $\times$ 192	2.3 weeks
Shanghai Academy of AI for Science	FuXi			70	4.69B	Not reported	Not reported
Shanghai AI Laboratory	FengWu			69	0.427B	Nvidia A100 $\times$ 32	17 days
Microsoft	Aurora			84	1.3B	Nvidia A100 $\times$ 32	2.5 weeks
ECMWF	AIFS	0.25 ∘ (Reduced Gaussian)	1110K	90 state vars	0.229B	Nvidia A100 $\times$ 64	1 week
IUM, CMA	StretchCast	0.875 ∘ Center Face	7.7K	69	0.083B	Hygon K100AI DCUs ( $\sim$ 35% A100 computing power) $\times$ 8	18 hours

Table 1: Comparison of training compute resources, number of parameters, and training time across different atmospheric models.

	Base	Base w/ SwinT FaceMixer	Base w/o FaceMixer	Base w/o GlobalMixer
Center RMSE $\downarrow$ 6h	0.115	0.123	0.127	0.125
Center RMSE $\downarrow$ 36h	0.210	0.222	0.231	0.230
Global RMSE $\downarrow$ 6h	0.095	0.102	0.106	0.104
Global RMSE $\downarrow$ 36h	0.189	0.201	0.212	0.209

Table 2: Overall ablation results at 6 h and 36 h. The full SCS_Base model achieves the best mean RMSE and ACC both over the center-refined face and globally, indicating that both the FaceMixer and the GlobalMixer make stable positive contributions to forecast quality.

Variable @ Lead	Base RMSE $\downarrow$	Base ACC $\uparrow$	Base w/ SwinT FaceMixer RMSE $\downarrow$	Base w/ SwinT FaceMixer ACC $\uparrow$	Base w/o FaceMixer RMSE $\downarrow$	Base w/o FaceMixer ACC $\uparrow$	Base w/o GlobalMixer RMSE $\downarrow$	Base w/o GlobalMixer ACC $\uparrow$
Q850 [kg kg -1 ] @ 6h	0.001	0.985	0.001 (+6.0%)	0.983 (-0.2%)	0.001 (+9.5%)	0.982 (-0.3%)	0.001 (+8.0%)	0.982 (-0.3%)
Q850 [kg kg -1 ] @ 36h	0.001	0.955	0.001 (+6.1%)	0.951 (-0.5%)	0.001 (+10.4%)	0.946 (-1.0%)	0.001 (+10.4%)	0.946 (-1.0%)
T2m [K] @ 6h	1.113	0.994	1.233 (+10.7%)	0.992 (-0.1%)	1.269 (+14.0%)	0.992 (-0.2%)	1.329 (+19.3%)	0.991 (-0.3%)
T2m [K] @ 36h	1.578	0.988	1.720 (+9.0%)	0.986 (-0.2%)	1.798 (+13.9%)	0.984 (-0.4%)	2.184 (+38.4%)	0.977 (-1.1%)
T850 [K] @ 6h	0.666	0.997	0.724 (+8.7%)	0.996 (-0.1%)	0.750 (+12.6%)	0.996 (-0.1%)	0.756 (+13.4%)	0.996 (-0.1%)
T850 [K] @ 36h	1.062	0.992	1.145 (+7.8%)	0.991 (-0.1%)	1.220 (+14.9%)	0.990 (-0.2%)	1.278 (+20.4%)	0.989 (-0.3%)

Table 3: Ablation results for key variables on the center-refined face at 6 h and 36 h. Consistent with the global key-variable results reported in Supplementary Table S1, the full SCS_Base model remains the best-performing variant over the center region.

Variable @ Lead	Base RMSE $\downarrow$	Base ACC $\uparrow$	Base + SwinT FaceMixer RMSE $\downarrow$	Base + SwinT FaceMixer ACC $\uparrow$	Base w/o FaceMixer RMSE $\downarrow$	Base w/o FaceMixer ACC $\uparrow$	Base w/o GlobalMixer RMSE $\downarrow$	Base w/o GlobalMixer ACC $\uparrow$
Q850 [kg kg -1 ] @ 6 h	0.001	0.979	0.001 (+6.2%)	0.976 (-0.3%)	0.001 (+10.2%)	0.974 (-0.5%)	0.001 (+9.0%)	0.975 (-0.4%)
Q850 [kg kg -1 ] @ 36 h	0.001	0.925	0.001 (+5.7%)	0.917 (-0.9%)	0.001 (+10.9%)	0.908 (-1.9%)	0.001 (+10.3%)	0.909 (-1.7%)
T2m [K] @ 6 h	0.988	0.993	1.096 (+10.9%)	0.991 (-0.2%)	1.139 (+15.3%)	0.991 (-0.2%)	1.204 (+21.9%)	0.990 (-0.3%)
T2m [K] @ 36 h	1.490	0.984	1.616 (+8.5%)	0.982 (-0.3%)	1.696 (+13.8%)	0.980 (-0.5%)	2.017 (+35.4%)	0.972 (-1.3%)
T850 [K] @ 6 h	0.639	0.996	0.698 (+9.3%)	0.995 (-0.1%)	0.728 (+14.1%)	0.995 (-0.1%)	0.724 (+13.3%)	0.995 (-0.1%)
T850 [K] @ 36 h	1.102	0.988	1.190 (+7.9%)	0.986 (-0.2%)	1.274 (+15.6%)	0.984 (-0.4%)	1.293 (+17.3%)	0.984 (-0.5%)

Table S1: Global key-variable ablation results at 6 h and 36 h. This supplementary table shows that the full SCS_Base model remains the best-performing variant on the same representative variables under global evaluation.

Explore similar work

AirCast-SR: A Foundation Model for Kilometer-Scale Atmospheric Super-Resolution via Latent Consistency Diffusion
May 20, 2026Somnath Luitel, Manmeet Singh, Joshua Durkee +11Geospatial Foundation Models Spatiotemporal Forecasting
Department of Earth, Environmental, and Atmospheric Sciences, Western Kentucky University, Bowling Green, KY, USA · NASA Goddard Space Flight Center, Greenbelt, MD, USA · The University of Texas at Austin, Austin, TX, USA +3
U-Cast: A Surprisingly Simple and Efficient Frontier Probabilistic AI Weather Forecaster
Apr 10, 2026Salva Rühling Cachay, Duncan Watson-Parris, Rose YuWeather Forecasting Probabilistic Forecasting
UC San Diego
WeatherNext 3: Increasing resolution and performance of global weather models with raw observations
Sep 3, 2026Stephan Rasp, Boris Babenko, Dominic Masters +22Probabilistic Forecasting Weather Forecasting
1Google DeepMind · 2Google Research · 3Google

May 20, 2026J/K move · Enter open · S save

AirCast-SR: A Foundation Model for Kilometer-Scale Atmospheric Super-Resolution via Latent Consistency Diffusion

Somnath Luitel, Manmeet Singh, Joshua Durkee, Abdullah Al Fahad+10

Department of Earth, Environmental, and Atmospheric Sciences, Western Kentucky University, Bowling Green, KY, USA · NASA Goddard Space Flight Center, Greenbelt, MD, USA · The University of Texas at Austin, Austin, TX, USA+3

Geospatial Foundation Models Spatiotemporal Forecasting Diffusion Models Weather Forecasting

Open paper AskPDF arXiv

Operational weather prediction at kilometer scales remains computationally prohibitive for traditional numerical weather prediction (NWP) models, limiting forecast access for applications in energy, agriculture, and disaster management that require fine-grained spatiotemporal detail. Here we introduce AirCast-SR, a foundation model for atmospheric super-resolution that downscales global AI weather forecasts from 0.25 degree (~28 km) to 1 km horizontal resolution at hourly temporal resolution, producing 67-hour forecasts of eight coupled surface variables simultaneously. EarthMind-SR employs a three-dimensional U-Net conditioned within a Latent Consistency Model (LCM) diffusion framework, trained on patch-based samples over the contiguous United States (CONUS) using GraphCast forecasts as input and NOAA's Analysis of Record for Calibration (AORC) as the target. The model achieves near-zero bias across all variables and lead times, and its radial power spectral density analysis demonstrates preservation of fine-scale atmospheric structure at wavelengths of 10 km to 100 km where coarser models lose spectral power. We validate EarthMind-SR across three CONUS case studies spanning winter, summer, and spring seasons, and demonstrate zero-shot global transferability over India and Germany using independent surface station observations without any retraining or fine-tuning. As an open-weights foundation model, EarthMind-SR establishes a new paradigm for kilometer-scale AI weather prediction and provides a platform for regional fine-tuning, distillation, and downstream applications in climate services and hazard forecasting.

Abstract

Explore similar work

AirCast-SR: A Foundation Model for Kilometer-Scale Atmospheric Super-Resolution via Latent Consistency Diffusion

U-Cast: A Surprisingly Simple and Efficient Frontier Probabilistic AI Weather Forecaster

WeatherNext 3: Increasing resolution and performance of global weather models with raw observations