Research map / Learning theory and algorithms / Prediction and decision losses
Towards a Self-contained Data-driven Global Weather Forecasting Framework
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.
Couples the FengWu AI forecasting model with an AI-embedded 4DVar data assimilation algorithm (model-error-aware objective, spherical-harmonic horizontal background covariance, autodiff solver) to run a self-contained global forecast system for a year using simulated ERA5 observations.
Problem
Data-driven weather forecasting models depend on analysis states produced by expensive traditional data assimilation such as 4DVar, which is computationally costly and inconsistent with AI models.
Approach
Replace the physics model in the 4DVar objective with the AI forecast model (1-hour variant), add a model-error covariance term Q to the observation covariance, model horizontal background error via spherical convolution using SHT, and solve for the analysis with autodiff and L-BFGS instead of hand-coded adjoints.
Key results
- FengWu-4DVar sustains stable cyclic assimilation and forecasting for one year, with analysis RMSE for z500 below the IFS 6-hour forecast on most variables.
- With 15% observation proportion, analysis accuracy is comparable to the IFS 6-hour forecast; 5% observations still converge (RMSE about 40 m2/s2 for z500).
- Assimilating a 6-hour window takes 29.3 s on one A100 GPU, roughly 100 times faster than traditional 4DVar on 256 processors.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use AI-embedded 4DVar instead of 3DVar | FengWu-4DVar · z500 analysis, 15% observation proportion | RMSE | Helped |
| Add the model error term to the objective function (ablation) | AI-embedded 4DVar · z500 analysis, cyclic forecasting | RMSE: above 200 m2/s2 after a week | Helped |
| Use spherical convolution instead of ordinary convolution to approximate horizontal correlation (ablation) | AI-embedded 4DVar · z500 analysis increment | analysis increment continuity | Helped |
| Add horizontal correlation to the background covariance approximation (ablation) | AI-embedded 4DVar · z500 analysis increment | analysis increment continuity | Helped |
| Reduce observation proportion from 15% to 5% | FengWu-4DVar · z500 analysis | RMSE: about 40 m2/s2 | Hurt |
| Increase observation noise standard deviation from 0.001 to 0.01 | FengWu-4DVar · z500 analysis | RMSE: from around 20 m2/s2 to over 70 m2/s2 | Hurt |
| Use AI-embedded 4DVar rather than 3DVar | FengWu-4DVar · z500 analysis with noisier observations | RMSE | Helped |
| Use different initial states | FengWu-4DVar · z500 analysis | RMSE: around 20 days | No clear effect |
| Remove the model error term from the objective function (ablation) | AI-embedded 4DVar · z500 analysis, cyclic forecasting | RMSE: above 200 m2/s2 after a week | Hurt |
Surprising findings
- Removing the model error term causes z500 analysis RMSE to rise above 200 m2/s2 within a week of cyclic forecasting.
Limitations
- Experiments use simulated ERA5-based observations rather than real-world observations.
- No end-to-end comparison with an operational 4DVar system.
- Vertical correlation of background error is not modeled.
Experimental setup
Datasets: ERA5
Benchmarks: ERA5 simulated observations
Metrics: RMSE Bias
Baselines: 3DVar IFS 6-hour forecast traditional 4DVar ordinary convolution horizontal correlation diagonal background covariance
Builds on: 4DVar FengWu Geer (2021) AI-embedded 4DVar torch-harmonics
Closest papers by meaning
- Tensor-Var: Efficient Four-Dimensional Variational Data Assimilation (ICML 2025 · similarity 0.87)
- DAWP: A framework for global observation forecasting via Data Assimilation and Weather Prediction in satellite observation space (NeurIPS 2025 · similarity 0.86)
- DiffDA: a Diffusion model for weather-scale Data Assimilation (ICML 2024 · similarity 0.86)
- Fixing the Double Penalty in Data-Driven Weather Forecasting Through a Modified Spherical Harmonic Loss Function (ICML 2025 · similarity 0.85)
- FNP: Fourier Neural Processes for Arbitrary-Resolution Data Assimilation (NeurIPS 2024 · similarity 0.84)
- Probablistic Emulation of a Global Climate Model with Spherical DYffusion (NeurIPS 2024 · similarity 0.84)
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.