WeatherNext 2: Google unveils its most advanced AI weather model yet
Google has launched WeatherNext 2, its fastest and most accurate AI weather forecasting model yet. It delivers hour-level global forecasts and can generate hundreds of possible scenarios in under a minute.
New Delhi: Google DeepMind and Google Research have introduced an AI-based model of weather forecasting called WeatherNext 2 that will provide improved forecasting with increased speed, accuracy, and resolution of weather systems on a global scale. The new system will deliver eight times quicker predictions and resolve within an hour, which is a significant improvement over the old-fashioned physics-based models that take hours on supercomputers. As weather can affect flight operations and even their day-to-day commuting, Google argues that the new model will enable agencies and users to make plans with certainty.
WeatherNext 2 is currently leaving the research laboratories and coming into actual practice. It already adds its forecast data to Earth Engine, Google Search, Gemini, Pixel Weather and Google Maps Platform through its Weather API. Google Cloud Vertex AI also has an early access program that will enable developers to run custom model inference to expand real applications across industries.
Predicting hundreds of possible scenarios
The new system has the capability of producing hundreds of weather results on a single input with independently trained neural networks and noise injection control. This allows real-world variability and allows worst-case scenarios to be captured, which is important information in climate-sensitive planning. The individual predictions require less than a minute on an individual TPU.
Higher accauracy across nearly all variables
WeatherNext 2 has a better result on 99.9 percent more variables of the atmosphere and forecast lead times of 15 days compared with the old WeatherNext model. It is based on a new Functional Generative Network (FGN) method, which adds noise to the architecture to retain physically consistent predictions. The model can predict multifaceted joints such as heat-affected areas or energy production in a wind farm, yet all it has been trained to do are temperature and humidity (so-called marginals).
According to Google, WeatherNext 2 marks the transition in the field from working on AI experiments to large-scale use. The company is investigating new sources of data, wider access and sustained performance enhancement. Using open data tools and refined modelling, Google is committed to assisting global researchers, developers and businesses to address weather-based challenges and construct resilient systems in the future.