Artificial intelligence aids in predicting sudden stratospheric warming events

The polar vortex, a mass of cold air that rotates around the North Pole, can be disrupted by sudden stratospheric warming (SSW) events. These events release cold air waves that move further south than usual, causing sharp temperature drops in regions such as North America and Europe. However, these events are difficult to predict, making it challenging to issue early warnings ^2,3.

Machine learning model

To improve the prediction of SSW, researchers developed a machine learning model. The model was trained based on the Holton-Mass model of stratospheric variability, which simulates two meta-stable regimes of the polar vortex: a strong and a weak one. The transition from the strong regime to the weak one is what causes SSW events ^1,2,3.

The model, built with a conditional variational autoencoder, was able to reproduce both the strong and weak regimes of the polar vortex. Additionally, the researchers performed a principal component analysis to understand how the model represents these characteristics. They identified four physically interpretable clusters, corresponding to the strong and weak states, as well as the transition trajectories between them ^1,2,3.

Within these clusters, the researchers found an interpretable signal that can predict whether a disruption in the polar vortex is about to occur. This suggests that the machine learning model can be repurposed to predict future SSW events in the real world. Furthermore, the work demonstrates that complex atmospheric features can be modeled and understood using artificial intelligence ^1,2,3.