Weather predictions have to capture the total range of possibilities — including worst case scenarios, that are crucial to plan for.
WeatherNext 2 can predict lots of of possible weather outcomes from a single place to begin. Each prediction takes lower than a minute on a single TPU; it might take hours on a supercomputer using physics-based models.
Our model can be highly skillful and able to higher-resolution predictions, all the way down to the hour. Overall, WeatherNext 2 surpasses our previous state-of-the-art WeatherNext model 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 brand new AI modelling approach called a Functional Generative Network (FGN), which injects ‘noise’ directly into the model architecture so the forecasts it generates remain physically realistic and interconnected.
This approach is especially useful for predicting what meteorologists seek advice from as “marginals” and “joints.” Marginals are individual, standalone weather elements: the precise temperature at a selected location, the wind speed at a certain altitude or the humidity. What’s novel about our approach is that the model is barely trained on these marginals. Yet, from that training, it learns to skillfully forecast ‘joints’ — large, complex, interconnected systems that rely upon how all those individual pieces fit together. This ‘joint’ forecasting is required for our most useful predictions, equivalent to identifying entire regions affected by high heat, or expected power output across a wind farm.
