Hierarchy of extreme-event predictability in turbulence revealed by machine learning

Yuxuan Yang, Chenyu Dong, Gianmarco Mengaldo

TL;DR

Researchers have used generative AI (diffusion models) to discover a "predictability hierarchy" in turbulent flows, revealing that some extreme events can be foreseen four times longer than others. By analyzing these events, they found that the persistence of large-scale, four-vortex "quadrupole" structures—rather than small-scale chaos—is the primary driver of how far into the future we can accurately predict a disaster.

What Problem Does This Paper Address?

Turbulence is the quintessential chaotic system. From atmospheric weather to industrial pipelines, turbulent flows are defined by "extreme events"—rare, high-impact surges in energy or enstrophy (a measure of rotational intensity).

Traditionally, scientists quantify predictability using the "Lyapunov time," which measures how fast a tiny error in initial conditions grows. However, this is a global average. In reality, predictability is state-dependent: some moments in a flow are inherently more stable and predictable than others. Quantifying this "event-wise" predictability usually requires knowing the exact governing equations (Navier-Stokes) and running massive, expensive "ensembles" (thousands of parallel simulations). This paper asks: Can we diagnose these predictability limits directly from data, without the equations, and find out why some extremes are easier to spot than others?

Methods

The authors employ a multi-step data-driven framework:

  1. Surrogate Modeling: They trained an autoregressive conditional diffusion model on fluid dynamics data. Unlike standard models that give a single "best guess," diffusion models learn the entire probability distribution of future states, allowing the researchers to generate "synthetic ensembles" very cheaply.
  2. Skill Scoring: They used the Continuous Ranked Probability Score (CRPS) to compare the AI’s forecast against a "climatological" baseline (a guess based on long-term averages).
  3. Predictability Horizon ($T^*_e$): They defined the horizon as the specific lead time where the AI's forecast skill finally drops to the level of a random guess.
  4. Spectral Filtering & Coherent Structure Analysis: To find the physical cause of predictability, they systematically stripped away different scales of the flow (from small eddies to large waves) and used the Q-criterion to identify and track vortex structures.

Datasets

The study uses data from Direct Numerical Simulations (DNS) of two-dimensional Kolmogorov flow at a Reynolds number of 100.

  • Grid: $128 \times 128$ spatial resolution.
  • Events: The test set focused on 211 extreme events, defined as the top 1% of the enstrophy distribution.
  • Validation: The AI-generated predictability limits were validated against traditional, computationally expensive DNS perturbation ensembles to ensure the "machine-learned" physics were accurate.

Key Insights and Ideas

  • The Predictability Hierarchy: Predictability isn't a single number. The authors found a massive spread: some extremes lose predictability in just 1 Lyapunov time, while others remain predictable for over 4.
  • Large Scales Rule: By "filtering" the data, the authors found that predictability is almost entirely governed by large-scale structures. Removing small-scale fluctuations (eddies smaller than ~30% of the domain size) had virtually no impact on the predictability horizon.
  • The "Quadrupole" Precursor: Extreme events are preceded by intense "strain cores" that organize four vortices into a quadrupolar packet.
  • Persistence is Key: The "aha moment" was discovering that the lifetime of these quadrupolar structures is the "smoking gun." If the quadrupole persists longer in a Lagrangian frame (moving with the flow), the resulting extreme event is significantly more predictable.

Key Results

  • Quantified Horizons: The AI successfully categorized events into high-, medium-, and low-predictability regimes.
  • Critical Scale: They identified a critical physical wavelength ($\lambda_c \approx 0.3L$) below which spatial details do not contribute to the predictability of the extreme event.
  • Structural Link: High-predictability events showed a statistically significant correlation with longer-lived quadrupolar precursors (confirmed via a bootstrap test with $p = 0.017$).
  • Equation-Free Validation: The diffusion model’s estimates qualitatively matched the divergence behavior seen in "ground truth" DNS ensembles, proving that ML can act as a reliable surrogate for complex fluid physics.

Limitations and Future Work

  • Dimensionality: The study is limited to 2D Kolmogorov flow. While a standard benchmark, real-world turbulence is 3D, where vortex stretching (a mechanism absent in 2D) plays a major role.
  • Reynolds Number: The Reynolds number ($Re=100$) is relatively low. Higher turbulence levels might introduce more complex scale interactions that could challenge the "large-scale dominance" found here.
  • Observational Data: While the method is "equation-free," it still requires high-quality, time-resolved data. Applying this to sparse or noisy experimental/satellite observations is the next major hurdle.
  • Future Directions: The authors suggest this framework could be used to provide "early warning" systems for atmospheric extremes or to optimize control strategies in engineering flows.

Summary updated

Abstract

Extreme-event predictability in turbulence is strongly state dependent, yet event-by-event predictability horizons are difficult to quantify without access to governing equations or costly perturbation ensembles. Here we train an autoregressive conditional diffusion model on direct numerical simulations of the two-dimensional Kolmogorov flow and use a CRPS-based skill score to define an event-wise predictability horizon. Enstrophy extremes exhibit a pronounced hierarchy: forecast skill persists from $\approx 1$ to $> 4$ Lyapunov times across events. Spectral filtering shows that these horizons are controlled predominantly by large-scale structures. Extremes are preceded by intense strain cores organizing quadrupolar vortex packets, whose lifetime sharply separates long- from short-horizon events. These results identify coherent-structure persistence as a governing mechanism for the predictability of turbulence extremes and provide a data-driven route to diagnose predictability limits from observations.
BibTeX
@misc{yang2026hierarchy,
  author = {Yuxuan Yang and Chenyu Dong and Gianmarco Mengaldo},
  title = {{Hierarchy of extreme-event predictability in turbulence revealed by machine learning}},
  year = {2026},
  eprint = {2603.13789},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2603.13789},
}
Figure