Google's new WeatherNext 2: AI weather forecasts are 8 times faster and support generating hundreds of scenarios.

Google DeepMind and Google Research have announced the launch of a brand new AI weather forecasting model, WeatherNext 2, which emphasizes that this model can enhance the speed of global weather forecasts by 8 times and achieve high-resolution forecasts down to the “1-hour level.” More importantly, it can generate hundreds of possible weather scenarios from a single initial data point, allowing users, meteorological agencies, and industries to better grasp the potential range of extreme weather. Currently, WeatherNext 2 has been integrated into several applications, including Google Maps and Gemini.

Both speed and resolution have been significantly improved, and AI forecasting has been fully upgraded.

The official statement indicates that WeatherNext 2 can run a forecast on Google's own TPU in less than 1 minute, which is much more efficient than the several hours required by traditional physical models. It can also output forecasts at a higher resolution, detailing changes down to “hourly” variations, which aids in flight scheduling, supply chain management, and even the commuting and daily arrangements of the general public.

The officials pointed out that WeatherNext 2 outperformed the previous generation WeatherNext in 99.9% of weather variables and in forecast times of 0 to 15 days, including multiple indicators such as temperature, wind speed, and humidity, with a significant overall improvement in accuracy.

The image is a flowchart of the WeatherNext 2 weather scenario simulation process, with core technology FGN, creating a more realistic weather simulation.

The performance breakthrough of WeatherNext 2 is primarily due to the adoption of a new model architecture called Functional Generative Network (FGN). This architecture injects “noise” (Noise) into the model, allowing it to maintain natural variation while generating weather scenarios, without deviating from physical laws.

In order to make the forecast results diverse yet consistent, the system uses multiple independently trained neural networks, which then generate a series of coherent scenario versions through noise adjustment. Google also specifically mentioned that this model is trained using a “single variable” (Marginals), such as temperature or wind speed, but is able to predict the “linked systems” (Joints), such as the impact of heat waves on an entire region, or the overall distribution of power generation from the wind field.

The ability to infer complex systems from simple training data is the core advantage of WeatherNext 2.

( Note: The noise here refers to the random variations deliberately introduced into the AI model, allowing for a bit more randomness in the model's calculations to produce different but still physically plausible weather changes. The linked system refers to the overall weather system formed by the interaction of multiple weather variables, rather than single-point data. )

Generate hundreds of scenarios at once, making extreme weather predictions more grounded.

WeatherNext 2 can generate hundreds of weather results from the same set of input data in less than 1 minute, while keeping each one physically reasonable and internally consistent.

This ability is particularly important for “high uncertainty events” such as typhoons, heavy rain, and heat waves. Google has also utilized this technology to support meteorological agencies in experimental typhoon forecasting, helping to assess the worst-case scenarios and possible ranges of change.

Fully introduce Google products and open up to the global community.

Currently, Google has officially integrated WeatherNext 2 into multiple client and enterprise products. General users can see the new forecasting system in Google Search, Gemini, Pixel Weather, and the Weather API of Google Maps Platform, and it will also be added to Google Maps in the future.

Developers can obtain forecast data through Earth Engine and BigQuery, and use their own AI weather models on Google Cloud's Vertex AI.

Google promotes global research, introduces more data and expands its usage.

Google stated that it will continue to research how to integrate more data sources in the future to further enhance model capabilities, and will also continue to open new tools for researchers, developers, and businesses to use whenever possible.

The officials hope to leverage powerful AI models and open data to assist the global community in making important decisions regarding energy, transportation, and climate more quickly, and to drive more scientific breakthroughs.

This article discusses Google's newly launched WeatherNext 2: AI weather forecasting speeds up by 8 times, supporting the generation of hundreds of scenarios, first appearing in Chain News ABMedia.