This paper builds an artificial intelligence emulator that can recreate rare breakdowns of the polar vortex in a simple, idealized model of the stratosphere. The target system is the stochastic Holton–Mass model, a 75‑dimensional physical model that captures two long‑lived states — a strong vortex and a weak vortex — and occasional random jumps between them that are qualitatively like sudden stratospheric warming (SSW) events. Those jumps are rare, so they are hard to learn with data‑driven methods because most training data sit in the stable states.

The authors train a probabilistic deep learning model called a Conditional Variational Autoencoder (CVAE). This CVAE is ResNet‑inspired, with six‑layer encoder and decoder networks and explicit conditioning on the current state. The emulator is trained to produce the distribution of possible system states one day ahead rather than a single forecast. The internal representation of the model uses a 32‑dimensional latent space that the network learns during training.

At a high level, the CVAE works by turning the current state into a compact internal code and then sampling from that code to generate plausible next states. Because it is probabilistic, the emulator can represent uncertainty and the range of possible outcomes that arise from the model’s random forcing. The paper reports that the emulator reproduces many diagnostics of the Holton–Mass model: short‑term dynamics, long‑term steady distributions, how long the system tends to stay in each regime (regime persistence), the rate of rare transitions, the transition committor function (the probability that a given state will soon jump to the other regime), and the expected lead time before a transition.

Beyond emulation accuracy, the authors analyze what the CVAE learned inside. They apply principal component analysis (PCA) to the 32‑dimensional latent space and find a clear, unsupervised separation into four clusters. These clusters map to physically meaningful situations: strong versus weak vortex, and within each of those, states that are either stable or prone to transition. In other words, the emulator not only reproduces the model behavior but also organizes its internal code into interpretable patterns that signal when a transition is more likely.