WeatherNext 2: The Impact of Google’s AI Forecasting Model

Weather forecasting has become a proving ground for AI – with machine learning (ML) models increasingly challenging traditional physics-based approaches that have dominated meteorology for decades. 

The stakes are high, affecting everything from international logistics to personal travel plans. 

In response, Google DeepMind and Google Research have now launched WeatherNext 2, a model that processes predictions eight times faster than its predecessor while offering resolution down to one-hour intervals.

The WeatherNext team explains that “the weather affects important decisions we make everyday – from global supply chains and flight paths to your daily commute”. 

“Hundreds of millions of people turn to Google Search every week for the weather,” says Nick Fox, Senior Vice President (SVP) of Knowledge and Information at Google.

“We just launched our powerful WeatherNext AI Models in Google Search to help upgrade weather forecasts. This is our most advanced weather forecasting tech in Search to date.

“Our WeatherNext AI Models lead to big improvements in Search, like (1) more accurate weather forecasts further out – with significant improvements in the 2-10 day out range (2) a massive (!!) improvement in weather forecast granularity (in other words, highly-localised forecasts!).”

How WeatherNext 2 works

The system generates hundreds of weather scenarios from a single starting point, with each prediction requiring less than one minute on a Tensor Processing Unit (TPU). 

Traditional physics-based models running on supercomputers need hours to produce comparable results.

Google has already integrated the technology across its consumer products, including Search, the Gemini chatbot and Pixel Weather. 

The company has also made forecast data available through Earth Engine and BigQuery, its geospatial analysis and data warehouse platforms, while launching an early access programme on Google Cloud’s Vertex AI platform for custom model inference.

Why WeatherNext 2 uses a functional generative network approach

The model employs a technique called Functional Generative Network, which introduces noise directly into the architecture to produce multiple forecast scenarios. 

The team notes that “weather predictions need to capture the full range of possibilities – including worst case scenarios, which are the most important to plan for”.

This method trains the system on individual weather elements such as temperature, wind speed and humidity, which meteorologists term “marginals”. 

From this training, the model learns to forecast “joints”, the large interconnected weather systems that depend on how individual components interact. 

The team says this joint forecasting is “required for our most useful predictions, such as identifying entire regions affected by high heat, or expected power output across a wind farm”.

The architecture uses independently trained neural networks that inject noise in function space, creating variability in predictions while maintaining coherence across scenarios. 

Each network operates separately but produces interconnected results that reflect realistic atmospheric conditions.

Google states the model outperforms the original WeatherNext on 99.9% of variables and lead times spanning zero to 15 days. 

The system covers standard meteorological variables including temperature, wind and humidity across different pressure levels and timeframes.

How Google expands deployment across products and platforms

The company has deployed the technology to support weather agencies in making decisions based on multiple scenario analyses. 

The team notes it has used the technology in “experimental cyclone predictions”, working directly with meteorological services to test real-world applications.

Google Maps will incorporate WeatherNext 2-powered weather information to expand beyond current implementations in Search and mobile applications. 

The model’s hourly resolution is an increase in temporal detail compared to standard forecasting intervals, enabling more precise predictions for applications from supply chain management to flight path planning.

Google has not disclosed the size of the neural networks used in WeatherNext 2 or the volume of training data required to achieve the reported performance levels. 

Google operates data centres equipped with TPUs, which provide the computational infrastructure for both training and inference of large-scale ML models.

The WeatherNext team says it is “actively researching capabilities to improve our models, including integrating new data sources and expanding access even further”. 

The release also forms part of Google’s geospatial and AI initiatives, which include Earth Engine for planetary-scale environmental data analysis and AlphaEarth Foundations for earth observation.

Multiple technology companies and research institutions have developed ML models for atmospheric prediction in recent years, with approaches varying in architecture and training methods as AI systems achieve performance gains over traditional numerical weather prediction.

The team adds: “By providing powerful tools and open data, we hope to accelerate scientific discovery and empower a global ecosystem of researchers, developers and businesses to make decisions on today’s most complex problems and build for the future”.