Meteorology stops being just maps and symbols and becomes decisions that affect supply chains, flight operations, and the safety of communities. Google DeepMind and Google Research introduce WeatherNext 2, a weather-prediction model that promises speed, higher resolution, and a richer probabilistic view so you can make better decisions.
What WeatherNext 2 brings
WeatherNext 2 delivers forecasts up to 8x faster and offers temporal resolution down to 1 hour. What does that mean in practice? You can generate hundreds of scenarios from the same starting point, and each prediction takes less than a minute on a single TPU. For context: physics-based models can take hours on a supercomputer to reach similar results.
WeatherNext 2 outperforms the previous model on 99.9% of variables (temperature, wind, humidity) and across horizons from 0 to 15 days.
The model produces hundreds of possible outcomes per input, which is essential for planning not only the most likely scenario but also the extreme cases that demand immediate attention.
How it works (accessible technical explanation)
The central novelty is an architecture called Functional Generative Network (FGN). Instead of training a network for a single output, WeatherNext 2 uses neural networks trained independently and injects noise in function space. Simply put, noise is introduced inside the architecture to generate coherent, physically realistic variability across predictions.
This lets the model capture two important concepts meteorologists use:
Marginals: individual variables like temperature at a point, wind speed at a certain altitude, or humidity.
Joints: complex patterns that involve interactions between many variables, for example a region affected by a heatwave or the expected output of a wind farm.
What’s striking is that WeatherNext 2 is trained only on marginals and still learns to generate joints with high fidelity. To measure probabilistic quality it uses CRPS (Continuous Ranked Probability Score), where WeatherNext 2 improves the score compared to prior generation, indicating more reliable probabilistic forecasts.
Performance and scalability
Generates hundreds of scenarios per input.
Each prediction takes less than a minute on a TPU.
Temporal resolution down to 1 hour and coverage from 0 to 15 days.
Improvements across nearly all relevant meteorological variables.
Thanks to its efficiency, WeatherNext 2 is already being tested for cyclone prediction and as support for meteorological agencies making decisions under uncertainty.
Integration and access for developers and organizations
Google is moving this lab advance into products and platforms you can use:
Forecast data available in Earth Engine and BigQuery.
Early access program for custom inference in Vertex AI on Google Cloud.
Integration into consumer products: Search, Gemini, Pixel Weather and Maps Platform Weather API. In the coming weeks it will reach weather information in Google Maps.
If you manage critical infrastructure, logistics operations, or emergency services, this gives you tools to assess risks with multiple scenarios in minutes.
Practical examples
A wind-farm operator can use the joints to estimate expected production and plan maintenance or power purchases.
An airline can simulate hundreds of alternative trajectories ahead of a strong front and optimize routes to reduce delay risk.
A disaster-management team can prioritize areas for evacuation by showing not just the most likely forecast but the probability of extreme events.
Do you see the difference between knowing what’s most likely and being ready for what’s possible?
Limitations and next research steps
WeatherNext 2 is a major step forward, but it doesn’t fully replace physical models in every context. AI models depend on the quality and coverage of data; that’s why Google’s teams are researching how to integrate new observation sources and expand global access.
Collaboration with meteorological agencies is also key to validate, calibrate, and use these forecasts in real operations.
Final reflection
We’re at a point where AI is not just about speed or isolated accuracy, but about offering multiple coherent scenarios that enable better, safer decisions. WeatherNext 2 shows how an architecture designed for uncertainty can transform how we plan for weather.
