Weather predictions need to include the full range of possibilities – including worst-case scenarios, for which it is most important to plan.
WeatherNext 2 can predict hundreds of possible weather outcomes from a single starting point. Each prediction takes less than a minute on a TPU; Using physics-based models would take hours on a supercomputer.
Our model is highly efficient and capable of producing high-resolution predictions every hour. Overall, WeatherNext 2 outperforms our previous state-of-the-art WeatherNext models on 99.9% of variables (e.g. temperature, wind, humidity) and lead times (0-15 days), enabling more useful and accurate forecasts.
This improved performance is enabled by a new AI modeling approach called Functional Generative Networks (FGNs), which inject ‘noise’ directly into the model architecture so that the predictions it produces remain physically realistic and interconnected.
This approach is particularly useful for predicting what meteorologists refer to as “marginals” and “couples”. Frontiers are individual, standalone weather elements: the exact temperature at a specific location, wind speed or humidity at a certain altitude. What is unique about our approach is that the model is trained only on these margins. Yet, from that training, it learns to efficiently predict ‘joints’ – large, complex, interconnected systems that depend on how all those different pieces fit together. This ‘combined’ forecast is essential for our most useful predictions, such as identifying entire areas affected by high heat, or the expected power output of a wind farm.