DeepMind and Google Research launched WeatherNext 3, a model that outperforms supercomputers, Microsoft, Nvidia, ECMWF on temperature, wind, humidity forecasts

Research teams at Google DeepMind and Google Research have launched the third generation of the weather forecasting model WeatherNext 3, a step that represents a structural shift in the use of deep-learning techniques and transformer models to simulate the atmosphere with higher accuracy and unprecedented temporal frequency, surpassing conventional supercomputers that rely on solving complex, computationally expensive physical equations.

The new model outperforms standard accuracy tests administered through the Operational Weather Bench platform of Bright Band, leading the competition against AI models developed by Microsoft and Nvidia and the European Centre for Medium-range Weather Forecasts (ECMWF), and also surpasses government physical forecasting systems in the United States and Europe on temperature, wind speed and humidity metrics.

Simulating complex weather phenomena from raw observation data breaks the constraints of slow generation of numerical matrices.

WeatherNext 3 addresses three historic challenges that have faced AI forecasting models: the broad geographic scope of predictions, the low accuracy of rain forecasts, and reliance on pre-structured data from government centers. By increasing the number of operations by a factor of 2.4 over the previous version and re-tuning the decoder heads, the model delivers spatially accurate forecasts up to 5 kilometers for core variables, with a 60 % improvement in rain-fall assessments compared with WeatherNext 2, and provides hourly predictions instead of the usual six-hour update cycles.

The most prominent technical capability lies in ingesting observations and data received directly from satellites around the clock without waiting for processing on supercomputers, along with training the model to directly forecast readings from ground observation stations located at airports and critical sites, thereby linking the forecasting task directly to field reality.

Energy applications and transport chains benefit from reduced inference cost and increased local forecasting accuracy.

This shift practically changes operating calculations for vital sectors in the Gulf, Egypt and the Levant, where accurate forecasts of solar illumination hours, cloud cover density and wind speed are a decisive factor for the stability of electricity grids and large renewable energy projects, especially solar and wind farms. Hourly forecasts with 5-kilometer resolution also enable port, airport and logistics complex administrations in the region to reduce the risk of weather disruptions and sudden storms at low computational cost through cloud platforms compared with the expense of building and managing local supercomputing centers.

Google intends to integrate the model’s outputs directly into Search, Google Maps and the Gemini assistant, as well as make it available to developers and researchers through its cloud infrastructure, giving startups and regulators the ability to build specialized services for smart agriculture and urban planning using fast, low-cost APIs.