> Original repository: google-deepmind/weathernext > Trending data: +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. The ECMWF model runs at roughly 9 km global grid resolution, and each forecast takes hours to compute.
This method works, but it has two bottlenecks: extremely high computational cost, and limited skill in predicting certain extreme weather events (such as typhoon tracks).
In 2023, DeepMind released GraphCast — using graph neural networks for weather forecasting, completing a 10-day forecast within one minute, outperforming ECMWF's HRES model. In 2024, GenCast used diffusion models for ensemble forecasting, again surpassing 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 parameters:
- Resolution: 0.25° (~30 km)
- Training data: ERA5 reanalysis data + 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 needs only one diffusion-model run
- Faster extreme-event warnings: cyclone track forecasts drop from hours to minutes
- Accessibility for developing countries: no supercomputer needed — one GPU suffices
WN2's cyclone forecasting version ran in real time during the 2025 Atlantic hurricane season, publicly known as 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 using graph neural networks. The first generation, proving AI could do weather forecasting. 2. WeatherNext Gen (GenCast): diffusion-model ensemble forecasting. Deterministic models give one answer; ensemble models give multiple possible answers to quantify uncertainty. 3. WeatherNext 2 (WN2): the latest generation, unifying atmospheric forecasting and cyclone forecasting. The same algorithm handles both daily weather and typhoon tracks.
The direction of evolution is clear: from deterministic to probabilistic, from single-task to unified multi-task.
Data Access: More Than Running the Model
One of WeatherNext's most interesting design choices: it doesn't just let you run the model — it provides multiple data access channels.
If you don't want to run the model yourself (WN2's weights come in 4 files, each in the GB range), you can directly use:
This means WeatherNext is not just an open-source repository but a complete weather data infrastructure. You can use the data for research, applications, and visualization without touching the model itself.
Why AI Weather Forecasting Matters
Traditional numerical weather prediction works like this: observations → assimilation → solving physics equations → output. Every step requires heavy computation and domain expertise.
AI weather forecasting works like this: observations → assimilation → neural network forward pass → output. Training requires massive compute, but inference takes only seconds.
This brings not just speed, but 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 combines them into a joint distribution — a classic probability trick: when a joint distribution is hard to estimate directly, approximate it with a combination of marginals.
This design choice shows DeepMind isn't just treating AI as a replacement for physics models — it's redesigning the mathematical framework of weather forecasting through a probabilistic lens.
Lessons for AI Research
WeatherNext's technical lineage offers AI researchers three lessons:
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 model weights but in data channels like Google Cloud, Earth Engine, and OpenMeteo. 3. From deterministic to probabilistic is the inevitable path: GraphCast was deterministic, GenCast probabilistic, WN2 unified both. This trajectory is happening across many areas of AI.
A Question Worth Pondering
WeatherNext open-sourced model weights and code, but not the training data or training pipeline. This means: you can use WN2 for forecasting, but you cannot reproduce WN2.
This is a common dilemma in AI for Science. Models can be open-sourced, but training data and compute cannot. Could this lead to AI weather forecasting being monopolized by a few big companies?
DeepMind's answer: open data channels (Google Cloud, OpenMeteo) so everyone can access forecast results. But that is not the same as being able to reproduce the model.
---
*Project: github.com/google-deepmind/weathernext* *Technical report: arXiv:2506.10772*