Weather forecasts must take into account the full range of possibilities – including worst-case scenarios, which are most important in planning.
WeatherNext 2 can predict hundreds of possible weather outcomes from a single starting point. Each prediction takes less than a minute on a single TPU; on a supercomputer using physics-based models, this would take hours.
Our model is also very useful and allows predictions at higher resolution, down to the hour. Overall, WeatherNext 2 outperforms 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 you to create more useful and accurate forecasts.
This improved performance is made possible by a new AI modeling approach called a Functional generative network (FGN), which injects “noise” directly into the model architecture so that the predictions it generates remain physically realistic and related.
This approach is particularly useful for predicting what meteorologists call “margins” and “ponds.” Marginalia are individual, independent weather elements: the exact temperature at a specific location, wind speed at a specific altitude, or humidity. What's new about our approach is that the model is trained only on these margins. However, with this training you learn to skillfully anticipate “connections” – large, complex, interconnected systems whose functioning depends on how all the individual elements fit together. This “joint” forecasting is required for our most useful predictions, such as identifying entire regions affected by high temperatures or the expected power output of a wind farm.
