Research map / Graph learning / Knowledge and temporal graphs
OneForecast: A Universal Framework for Global and Regional Weather Forecasting
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes OneForecast, a GNN-based global-regional nested weather forecasting framework with multi-scale region-refined mesh, gated multi-stream messaging, and a neural nested grid for regional high-resolution forecasts.
Problem
Data-driven weather forecasting struggles to balance global and high-resolution regional forecasts (boundary information loss), over-smooths extreme events and long-term rollouts, and lacks dynamic system modeling of multi-scale node-edge interactions.
Approach
A multi-scale icosahedral graph with region-refined mesh edges, a dynamic multi-head gated edge update plus multi-head node attention (multi-stream messaging), and a neural nested grid that feeds global model forecasts as trainable boundary/background forcing for regional models.
Key results
- On 1.5° WeatherBench2 global forecasting, OneForecast achieves 6-hour RMSE 0.0549 / ACC 0.9943 and 10-day RMSE 0.5918 / ACC 0.4457, best among Pangu, GraphCast, and FuXi retrained in the same framework.
- For 7-day ensemble-mean RMSE it reports 0.4393 (ENS) vs GraphCast ENS 0.4412, and 10-day ENS RMSE 0.5167 vs 0.5331 for GraphCast ENS.
- Ablation on regional forecasting: full NNG reaches RMSE 0.2180 / ACC 0.8856 versus 0.5828 / 0.2450 without NNG.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use OneForecast instead of GraphCast | Global weather forecasting · WeatherBench2 1.5°, 6-hour to 10-day | RMSE: 0.0626 -> 0.0549; 0.1304 -> 0.1231; 0.2861 -> 0.2732; 0.4597 -> 0.4468; 0.6009 -> 0.5918 | Helped |
| Use OneForecast instead of GraphCast | Global weather forecasting · WeatherBench2 1.5°, 6-hour to 10-day | ACC: 0.9928 -> 0.9943; 0.9705 -> 0.9737; 0.8705 -> 0.8825; 0.6692 -> 0.6888; 0.4275 -> 0.4457 | Helped |
| Add region-refined graph (ablation) | OneForecast · Region-refined data, 4-day | RMSE: 0.3793 -> 0.2609 | Helped |
| Add region-refined graph (ablation) | OneForecast · Region-refined data, 4-day | ACC: 0.6075 -> 0.8099 | Helped |
| Add neural nested grid instead of regional-only training (ablation) | OneForecast regional forecast · Regional forecast, 4-day | RMSE: 0.5828 -> 0.2180 | Helped |
| Add neural nested grid instead of regional-only training (ablation) | OneForecast regional forecast · Regional forecast, 4-day | ACC: 0.2450 -> 0.8856 | Helped |
| Use neural nested grid instead of boundary forcing (ablation) | OneForecast regional forecast · Regional forecast, 4-day | RMSE: 0.4428 -> 0.2180 | Helped |
| Use neural nested grid instead of boundary forcing (ablation) | OneForecast regional forecast · Regional forecast, 4-day | ACC: 0.4711 -> 0.8856 | Helped |
| Add multi-stream messaging instead of traditional MLP-based messaging (ablation) | OneForecast · WeatherBench2 1.5°, 4-day | RMSE: 0.3921 -> 0.2954 | Helped |
| Add multi-stream messaging instead of traditional MLP-based messaging (ablation) | OneForecast · WeatherBench2 1.5°, 4-day | ACC: 0.9305 -> 0.9577 | Helped |
| Add ensemble of 50 perturbed initial conditions | OneForecast · WeatherBench2, 7-day forecast | RMSE: 0.4268 -> 0.4393 | Hurt |
| Add ensemble of 50 perturbed initial conditions | OneForecast · WeatherBench2, 8-day forecast | RMSE: 0.4834 -> 0.4699 | Helped |
Limitations
- Global experiments use 1.5° resolution, which may limit cyclone prediction compared with 0.25° baselines, as the authors note.
- Theoretical high-pass claim rests on a heuristic gating assumption with loosely justified proportionality; not empirically validated beyond a spectral error plot.
Experimental setup
Datasets: WeatherBench2 ERA5
Benchmarks: WeatherBench2 ERA5 Best Track typhoon tracks (Yagi 2018, Molva 2020)
Metrics: RMSE ACC CSI SEDI
Baselines: Pangu-weather GraphCast FuXi FengWu Graph-EFM WeatherBench2 models
Builds on: GraphCast multigrid methods Graph-EFM Perlin noise ensemble perturbation
Closest papers by meaning
- Mesh Interpolation Graph Network for Dynamic and Spatially Irregular Global Weather Forecasting (NeurIPS 2025 · similarity 0.86)
- Scaling transformer neural networks for skillful and reliable medium-range weather forecasting (NeurIPS 2024 · similarity 0.85)
- Probabilistic Weather Forecasting with Hierarchical Graph Neural Networks (NeurIPS 2024 · similarity 0.85)
- HR-Extreme: A High-Resolution Dataset for Extreme Weather Forecasting (ICLR 2025 · similarity 0.84)
- OmniCast: A Masked Latent Diffusion Model for Weather Forecasting Across Time Scales (NeurIPS 2025 · similarity 0.84)
- ST-FiT: Inductive Spatial-Temporal Forecasting with Limited Training Data (AAAI 2025 · similarity 0.83)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.