New Machine Learning Method Helps Study Sudden Stratospheric Warming

The polar vortex is a swirling mass of cold air that circles the North Pole. Sometimes this swirling air is disrupted by sudden stratospheric warming (SSW) events. When this happens, waves of frigid air can travel farther south than usual, and temperatures across North America and Europe can drop sharply.

SSW events are still hard to predict. According to Phys.org, a team led by Boscu used a machine learning technique to learn how SSW events behave. The team learned this behavior from a physics-based simulation, which is a computer model built on the laws of physics.

The researchers trained their model on the Holton-Mass model of stratospheric variability. This model produces two metastable (stable for a while, but able to change) polar vortex regimes, one weak and one strong. A change from the strong regime to the weak regime is what causes SSW events.

To build their model, the authors used a conditional variational autoencoder. They then ran a principal component analysis (a method for finding the main patterns in data) to understand how the model represents features of the polar vortex. This analysis helped them find four physically interpretable clusters, which are groups of data that share common features. The clusters correspond to the strong and weak states and to the paths that move between them.

The authors say that a clear signal within these clusters can predict whether a polar vortex disruption is about to happen. They also suggest that the machine learning model could be repurposed to better predict future SSW events in the real world.

The work was published in the Journal of Geophysical Research: Machine Learning and Computation.