The one-sentence version
Traditional weather forecasting relies on numerical weather prediction: integrating atmospheric physics equations on supercomputers, where a single global forecast can take several hours. Since 2022, a wave of deep learning models (Huawei's Pangu-Weather, Google DeepMind's GraphCast and GenCast, NVIDIA's FourCastNet, etc.) trained only on decades of historical reanalysis data has matched or even surpassed the world's best numerical prediction systems across many medium-range forecast metrics—and can complete a forecast on a single accelerator card in minutes or even under a minute. In 2025, the European Centre for Medium-Range Weather Forecasts (ECMWF) deployed its own AI forecasting system, AIFS, into operational use, marking the transition of AI weather forecasting from research to practical operations.
Regarding Timeliness
This article is organized based on public information available at the time of writing and has not been cross-checked online item by item. This field is advancing extremely rapidly; please refer to the latest papers and official releases for specific metrics and rankings.
Layer 1: The Intuition
1.1 Two Forecasting Methods
| Numerical Weather Prediction | AI Weather Prediction | |
|---|---|---|
| Principle | Step-by-step calculation of future atmospheric states using physical equations (fluid dynamics, thermodynamics) | Learning "when the atmosphere looks like this, what it usually becomes in 6 hours" from decades of historical data |
| Computation | Supercomputers, several hours | Expensive during training; a few minutes on a single card during forecasting |
| Advantages | Grounded in physics; capable of handling unprecedented scenarios | Much faster; often has smaller errors in common scenarios |
| Dependencies | Observational data + physical models | Historical reanalysis data (which is itself a product of fusion between numerical models and observations) + current initial field |
1.2 Why is it so much faster?
To maintain stability, numerical prediction typically uses time steps of only a few minutes and must integrate step-by-step to reach ten days. AI models learn to "jump 6 hours at a time," reaching the tenth day in 40 autoregressive jumps, where each jump is merely a single neural network forward pass.
Layer 2: Representative Achievements
| Model | Institution / Time | Key points |
|---|---|---|
| FourCastNet | NVIDIA et al., 2022 | Based on Fourier Neural Operators; demonstrated the order-of-magnitude speed advantage of AI models |
| Pangu-Weather | Huawei, published in Nature in 2023 | 3D Earth-specific Transformer; one of the first AI models to overall surpass ECMWF's operational high-resolution deterministic forecasts in medium-range metrics |
| GraphCast | Google DeepMind, published in Science in 2023 | Graph Neural Network, 0.25° resolution; outperforms ECMWF's high-resolution deterministic forecasts on most evaluation targets; also has advantages in predicting tropical cyclone tracks and extreme weather |
| FengWu, Fuxi | Shanghai AI Lab, Fudan University, etc., 2023 | Global medium-range forecasting models from Chinese teams, extending the effective forecasting lead time |
| GenCast | Google DeepMind, published in Nature in late 2024 | Diffusion models for ensemble forecasting: generates a set of possible futures to provide probabilities; outperforms ECMWF's operational ensemble (ENS) on most metrics |
| AIFS | ECMWF, deployed operationally in 2025 | An AI forecasting system developed by the European Centre for Medium-Range Weather Forecasts, released operationally alongside its traditional numerical models |
| NeuralGCM | Google, 2024 | Hybrid model: physical dynamical core + neural network parameterization, capable of long-term climate simulation |
Layer 3: Significance, Limitations, and Future Directions
3.1 Why is this a "Revolution"
- Cost Reduction: Small meteorological departments, research teams, and developing countries can now run high-quality global forecasts;
- Cheaper Ensemble Forecasting: Models like GenCast can rapidly generate a large number of ensemble members, making probabilistic forecasting more reliable;
- Methodological Shift: For a long time, weather forecasting was considered a field where "physical equations must dominate"; the success of AI demonstrates that learning dynamics from data is feasible.
3.2 Limitations
| Limitation | Description |
|---|---|
| Dependency on reanalysis data | Training data itself comes from the fusion of numerical models and observations; the initial field still relies on traditional data assimilation systems |
| Extreme and unprecedented events | AI models may fail to extrapolate in situations that are rare or have never appeared in the training data; climate change is making "unprecedented" events increasingly common |
| "Smoothed" forecasts | To reduce mean error, deterministic AI models tend to provide blurry forecasts, underestimating extreme values; diffusion-based models have partially improved this |
| Physical consistency | Does not necessarily strictly conserve mass or energy |
| Interpretability | It is difficult for forecasters to understand "why" the model produces a certain forecast |
3.3 Directions
- End-to-End: Forecasting directly from raw observations (satellites, radar, ground stations), bypassing traditional data assimilation (research is already advancing this);
- Hybrid Models: Physical core + AI components (the approach of NeuralGCM), balancing stability and accuracy;
- Higher Resolution and Regional Forecasting: Short-term forecasting for severe convection and urban-scale weather;
- Climate Scale: Whether it can be used for decadal climate prediction remains an open question.
3.4 Connection to "World Models"
Weather forecasting can be viewed as a concrete, verifiable world model: given the current state, it predicts the next state, and the error can be objectively measured. Its success provides a powerful precedent for more general "world model" research (see Post-training of World Models): in domains where dynamics are sufficiently regular and data is sufficiently abundant, neural networks can indeed learn reliable state transitions.
Summary
| Layer | Key points |
|---|---|
| Change | AI models match or surpass top-tier numerical forecasts in medium-range forecasting, with speed advantages of several orders of magnitude |
| Representatives | Pangu, GraphCast, GenCast, FourCastNet, FengWu, Fuxi; ECMWF's AIFS has been operationalized |
| Limitations | Dependency on reanalysis and data assimilation; extreme events; smoothing; physical consistency |
| Directions | End-to-end, hybrid models, high resolution, climate scale |