Sudden stratospheric warming: an AI simulator reproduces rare shifts in the polar vortex and classifies them into four categories on its own

A conditional variational autoencoder replicates the Holton-Mass model, including the rare transitions between strong and weak vortices, and its latent space reveals a discernible physical structure.

Sudden stratospheric warming, which abruptly weakens the polar vortex and can cause cold waves at the surface, is difficult to model by machine learning: these events are rare, and therefore underrepresented in the data. A team led by C. Daniel Boscu and Justin Finkel is tackling this with a probabilistic emulator of the Holton-Mass stochastic model, a reference system that alternates between a strong and a weak vortex.

A latent space that "includes" regimes

The emulator, a conditional variational autoencoder inspired by ResNet, predicts the state of the system to the next day. It reproduces short-term dynamics, long-term distributions, the duration of each regime, the frequency of rare transitions and the average time to a switch.

More surprisingly, the principal component analysis of its latent space in 32 dimensions reveals, without supervision, four physically interpretable groups: strong or weak vortex, and stable or near-transition configurations. The authors see this as a way to improve early warning systems on extreme events.