30 km Global Weather Forecasting: How DeepMind's WeatherNext Rewrites Forecasting with AI
> Original repo: google-deepmind/weathernext > Trending stats: +105 stars today | Language: Python | Category: AI weather forecasting
A Problem Monopolized by Physics for 70 Years
Since the 1950s, weather forecasting has been the domain of physicists. The approach is straightforward: divide the atmosphere into a grid, write a set of fluid dynamics equations for each grid point, and solve them on supercomputers. ECMWF's model, with a global grid resolution of about 9 km, takes several hours to run per forecast.
This method works but has two bottlenecks: extremely high computational cost, and limited predictive skill for certain extreme events such as typhoon tracks.
In 2023, DeepMind released GraphCast — a graph neural network for weather forecasting that completes a 10-day forecast in under a minute with accuracy surpassing ECMWF's HRES model. In 2024, GenCast used diffusion models for ensemble forecasting, again beating traditional methods.
In 2025, WeatherNext 2 (WN2) arrived.
What Is WeatherNext 2
WN2 is a global medium-range atmospheric and cyclone forecasting model jointly developed by Google DeepMind and Google Research. Key specs:
- Resolution: 0.25° (~30 km)
- Training data: ERA5 reanalysis + ECMWF HRES operational data
- Training cutoff: 2024
- Forecast range: global, medium range (10–15 days)
- Special capabilities: 100-meter wind forecasting + tropical cyclone track forecasting
- Google Cloud: daily forecast data via Earth Engine, BigQuery, and Vertex AI
- WeatherLab: including cyclone track data
- OpenMeteo: API and interactive builder
- Ensemble forecasting becomes cheap: traditional ensembles run the model 50 times; AI runs one diffusion model
- Faster extreme-event warnings: cyclone track forecasts drop from hours to minutes
- Accessibility for developing countries: no supercomputer needed — one GPU suffices
The cyclone-focused version of WN2 ran in real time during the 2025 Atlantic hurricane season under the public name FNV3. The US National Hurricane Center (NHC) operates a post-processed version called GDMI.
Three Models, One Technical Lineage
The WeatherNext repository contains not just one model but a model family:
1. WeatherNext Graph (GraphCast): deterministic medium-range forecasting via graph neural networks. The first generation that proved AI could do weather forecasting. 2. WeatherNext Gen (GenCast): diffusion-based ensemble forecasting. Deterministic models give one answer; ensemble models give many possible answers to quantify uncertainty. 3. WeatherNext 2 (WN2): the newest generation, unifying atmospheric and cyclone forecasting. One algorithm handles both daily weather and typhoon tracks.
The evolutionary direction is clear: from deterministic to probabilistic, from single-task to multi-task unification.
Data Access: Not Just Running the Model
One of the most interesting design choices in WeatherNext is that it doesn't just let you run the model — it offers multiple data channels.
If you don't want to run the model yourself (WN2's weights span four files, each gigabytes in size), you can simply consume:
This makes WeatherNext not just an open-source repo but a full weather data infrastructure. You can build research, applications, and visualizations without ever touching the model itself.
Why AI Weather Forecasting Matters
Traditional numerical weather prediction pipeline: observations → assimilation → solving physics equations → output. Every step demands heavy computation and domain expertise.
AI weather forecasting pipeline: observations → assimilation → a forward pass through a neural network → output. Training is compute-heavy, but inference takes seconds.
Beyond raw speed, this unlocks new possibilities:
Key Details from the Technical Report
WN2's technical report (arXiv:2506.10772) is titled "Skillful joint probabilistic weather forecasting from marginals." The keyword is marginals.
Traditional ensemble forecasting runs the model many times to obtain a joint distribution. WN2 instead predicts each variable's marginal distribution first, then composes them into a joint distribution — a classic probability trick for when the joint distribution is hard to estimate directly.
This design choice shows that DeepMind isn't merely treating AI as a replacement for physics models, but is redesigning the mathematical framework of forecasting through a probabilistic lens.
Lessons for AI Research
The WeatherNext lineage offers three lessons for AI researchers:
1. Domain knowledge is irreplaceable: DeepMind's team includes meteorologists. Without their understanding of ERA5, HRES, and assimilation methods, a pure ML team couldn't have built this. 2. Infrastructure matters more than models: WN2's value lies not just in its weights but in the data channels — Google Cloud, Earth Engine, OpenMeteo. 3. From deterministic to probabilistic is an inevitable path: GraphCast was deterministic, GenCast probabilistic, WN2 unified both. This trajectory is playing out across many areas of AI.
A Question Worth Thinking About
WeatherNext open-sources the model weights and code — but not the training data or training pipeline. You can use WN2 to forecast, but you cannot reproduce WN2.
This is a common dilemma in AI for Science. Models can be open-sourced; training data and compute resources cannot. Could this lead to AI weather forecasting being monopolized by a few large companies?
DeepMind's answer: open data channels (Google Cloud, OpenMeteo) so everyone can access the forecasts. But that is not the same as being able to reproduce the model.
---
*Project: github.com/google-deepmind/weathernext* *Technical report: arXiv:2506.10772*