Turbulence spans scales. The atmosphere's energy cascades from 10,000-kilometer planetary waves down to millimeter-scale dissipation. No computer can resolve all scales simultaneously. So modelers truncate: simulate the large scales explicitly, model the small scales with a “closure” — a parameterization that represents what the unresolved scales would do if they were there.
Traditional closures are diffusive. They smooth. They average. They suppress exactly the extreme events — the hurricanes, the rogue waves, the sudden vortex breakdowns — that matter most. Five orders of magnitude of unresolved physics, compressed into a formula that throws away the tails.
Note (2026-11-12; read against the paper's full text): the paper's tests are five cases of two-dimensional turbulence with sinusoidal forcing in a periodic domain, a canonical prototype of atmospheric and oceanic flows, not the atmosphere; the abstract says up to five orders of magnitude fewer degrees of freedom, and the paper shows generalization to a flow at 15 times higher Reynolds number than its training data. Its baselines are the dynamic Smagorinsky and dynamic Leith closures, and it reports that the dynamic Leith one outperforms the Smagorinsky one in the cases it compares. The essay's picture of 10,000-kilometre planetary waves cascading to millimetres, its examples of hurricanes, rogue waves and vortex breakdowns, and its phrase 'five orders of magnitude of unresolved physics' (which mixes the degrees-of-freedom reduction with a range of scales) are not the paper's.
Scientific Multi-Agent Reinforcement Learning (arXiv:2603.03351) learns closures online using only the enstrophy spectrum as reward — a statistical fingerprint of how energy distributes across scales. The learned closure enables stable coarse simulations (five orders of magnitude fewer degrees of freedom) that reproduce high-fidelity statistics including the extremes.
The mechanism: the enstrophy spectrum captures the cascade. If the closure reproduces the cascade correctly, the large-scale statistics — including extreme events — emerge from the resolved dynamics. Traditional closures fail because diffusion destroys the cascade at the truncation scale. The learned closure preserves it because the RL agent is penalized when the spectrum departs from the target.
The structural insight: you don't need to predict extremes directly. You need to preserve the cascade that produces them. The extreme event is a consequence of multiscale interaction, not a separate phenomenon. Model the interaction, and the extremes appear for free.