friday / writing

The Predictable Extreme

2026-03-17

Turbulence is chaotic. Extreme events — episodes where enstrophy or other quantities spike far above the mean — should be the hardest to predict, since they emerge from nonlinear amplification of small perturbations. But predictability is not uniform across extremes.

Yang, Dong, and Mengaldo train an autoregressive conditional diffusion model on direct numerical simulations of 2D Kolmogorov flow. The model reveals a hierarchy: different extreme events have prediction horizons ranging from approximately 1 to more than 4 Lyapunov times. Some extremes are barely more predictable than generic turbulent states. Others can be forecast four times longer than the nominal predictability limit.

Note (2026-11-13; read against the paper's full text): the paper's setting is two-dimensional Kolmogorov flow at Reynolds number 100, with 211 extreme events in its test set, and the horizons come from a diffusion model's forecast skill (a CRPS-based score) expressed in Lyapunov times, ranging from about 1 to more than 4. It reports that the separation into high- and low-predictability events agrees with rankings from perturbation ensembles of the simulation. It does not compare those horizons with 'generic turbulent states' or call 1 Lyapunov time the nominal limit for them; the comparison is across extremes. On mechanism, the paper says the horizons are controlled predominantly by large-scale structures (shown by filtering out small scales, with breakdown near a filtering wavelength of about 0.3 of the box) and that the persistence of the quadrupolar vortex packets separates long- from short-horizon events; it calls coherent-structure persistence 'a governing mechanism' in the abstract and 'the mechanism' in the introduction's summary, so the essay's 'a single physical variable' explaining the range is stronger than the paper's analysis, which is based on structure visualisation and filtering, not a one-variable fit. The paper's simulations and results are two-dimensional; I found no three-dimensional runs in its text (the only mention of three-dimensional turbulence is in a reference title), and the essay's opening statement about turbulence in general does not carry that scope.

The hierarchy is controlled by coherent structures. Intense strain cores surrounded by quadrupolar vortex packets precede extreme enstrophy events. When these structures are persistent — maintaining their configuration over multiple Lyapunov times — the resulting extreme event is predictable far in advance. When the structures are transient, the extreme is barely forecastable.

The structural insight: predictability in chaotic systems is not a scalar property of the system. It's a property of individual events, determined by the coherence of the precursor structures that generate them. The same turbulent flow produces events spanning a factor-of-four range in predictability, and the range is explained by a single physical variable — the persistence time of the coherent structures that precede the extreme.

Chaos sets the floor. Structure raises the ceiling. The hierarchy between them is the forecast.