Research map / Time series / Long-term forecasting
Generalizing Weather Forecast to Fine-grained Temporal Scales via Physics-AI Hybrid Modeling
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 WeatherGFT, a physics-AI hybrid weather model that embeds a PDE kernel (300 s steps) with a learnable-router AI correction branch and multi-lead-time training, trained on hourly WeatherBench to forecast 30-minute precipitation beyond the dataset resolution.
Problem
Data-driven weather models are black-box mappings at fixed lead times and cannot forecast at time scales finer than the training dataset's temporal resolution (e.g., 30 minutes from hourly data).
Approach
A HybridBlock fuses a PDE-kernel physical evolution branch (finite-difference convolutions and integrals over atmospheric variables, 1/12 hour steps) with a parallel attention correction branch through a learnable router initialized at 0.5:0.5, plus a lead-time-conditional decoder and multi-lead-time training so the 2nd block's output can yield 30-minute forecasts.
Key results
- Generalized 30-min nowcast CSI@0.5 of 0.28 and CSI@1.5 of 0.17 for WeatherGFT, versus at most 0.26 and 0.13 for frame-interpolated baselines (FourCastNet/Keisler/ClimODE with Flavr or UPR).
- At 120-min lead time WeatherGFT reaches CSI@0.5 0.54 and CSI@1.5 0.40, the best among compared methods.
- Ablation on z500 RMSE at 1-h drops from 18.76 (attention baseline) to 16.66 with PDE kernel and multi-lead-time training.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use WeatherGFT instead of the strongest interpolation baseline | NASA precipitation nowcasting · 30-min generalized forecast | CSI@1.5: 0.13 -> 0.17 | Helped |
| Use WeatherGFT instead of the strongest interpolation baseline | NASA precipitation nowcasting · 90-min generalized forecast | CSI@1.5: 0.13 -> 0.16 | Helped |
| Use WeatherGFT instead of the strongest interpolation baseline | NASA precipitation nowcasting · 120-min forecast | CSI@1.5: 0.34 -> 0.40 | Helped |
| Add the PDE kernel to the Attention Block (ablation) | WeatherBench forecasting · 1-h | RMSE t2m: 0.52 -> 0.57 | Hurt |
| Add the PDE kernel to the Attention Block (ablation) | WeatherBench forecasting · 6-h | RMSE z500: 24.21 -> 21.78 | Helped |
| Add the PDE kernel to the Attention Block (ablation) | WeatherBench forecasting · 3-d | RMSE z500: 157.9 -> 153.8 | Helped |
| Add multi-lead time training to the PDE-kernel model (ablation) | WeatherGFT · 1-h | RMSE t2m: 0.57 -> 0.49 | Helped |
| Add multi-lead time training to the PDE-kernel model (ablation) | WeatherGFT · 6-h | RMSE z500: 21.78 -> 21.80 | No clear effect |
| Add multi-lead time training to the PDE-kernel model (ablation) | WeatherGFT · 3-d | RMSE z500: 153.8 -> 152.4 | Helped |
| Add the PDE kernel (ablation) | Attention Block baseline · 30-min nowcast | forecast availability | Helped |
Surprising findings
- The physical branch receives a consistently higher router weight than the AI branch, yet the AI weight grows with lead time as PDE errors accumulate.
- At 30-min, WeatherGFT has higher RMSE (0.72) than some interpolation baselines (e.g., 0.66-0.67), so it does not win on every metric.
Limitations
- Only five atmospheric PDEs are modeled, far from full atmospheric dynamics.
- Experiments are limited to 128x256 spatial resolution.
- Minimum evolution scale is 15 minutes, but 15-minute generalized predictions could not be evaluated due to lack of validation data.
Experimental setup
Datasets: WeatherBench ERA5 NASA satellite precipitation (test only)
Benchmarks: WeatherBench NASA satellite precipitation
Metrics: RMSE CSI@0.5 CSI@1.5 bias energy conservation
Baselines: FourCastNet Keisler ClimODE ECMWF-IFS Flavr UPR
Builds on: Swin Transformer FourCastNet ClimODE Pangu-Weather Earthformer
Closest papers by meaning
- Scaling transformer neural networks for skillful and reliable medium-range weather forecasting (NeurIPS 2024 · similarity 0.85)
- ChaosBench: A Multi-Channel, Physics-Based Benchmark for Subseasonal-to-Seasonal Climate Prediction (NeurIPS 2024 · similarity 0.84)
- Towards a General Time Series Forecasting Model with Unified Representation and Adaptive Transfer (ICML 2025 · similarity 0.84)
- Continuous Ensemble Weather Forecasting with Diffusion models (ICLR 2025 · similarity 0.84)
- Explanation of Dynamic Physical Field Predictions using WassersteinGrad: Application to Autoregressive Weather Forecasting (NeurIPS 2026 · similarity 0.83)
- OmniCast: A Masked Latent Diffusion Model for Weather Forecasting Across Time Scales (NeurIPS 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.