Brandstetter team's new paper: turbulence must be studied as a 4D space-time phenomenon

On October 26, Johannes Brandstetter's team released a new paper, "Let's Do Turbulence Right: In Space and Time!" (arXiv:2610.32601), drawing on 28TB of time-resolved DNS (direct numerical simulation) data to study how numerical choices affect spatiotemporal coupling dynamics in DNS records. The core conclusion: turbulence must be studied as a 4D (3+1 dimensional) spacetime phenomenon. The same day, he posted a long X thread aimed at the AI4Science community explaining and defending the argument—well worth AI researchers' attention.

Confirmed

• Core claim: numerical choices in simulations strongly shape turbulence properties; since the AI4Science community widely uses simulated flows to train and evaluate models, simulation quality directly determines model success or failure.
• Meaning of (3+1) dimensions: at each (x, y, z, t) point, velocity has three components; a single snapshot gives the volume field at one instant—only the "movie" reveals its temporal evolution.
• On timescales, the authors cite Tennekes' 1975 sweep hypothesis: large-scale motions "sweep" small eddies past an observation point faster than the eddies break up on their own; when sweeping dominates, correlation times scale roughly as 1/(Uk), accompanied by classic spectral plots of characteristic time τ versus wavenumber k (curves for τD, τR/τS, τC, etc.).
• On energy spectra, they present the famous mid-range scaling law E(k) ∝ k^(−5/3), derived by Kolmogorov in 1941 together with Obukhov's work.
• Competition within the Navier–Stokes equations: the convective term (u·∇)u describes the flow carrying and deforming itself, while viscosity smooths out velocity differences; with the Reynolds number Re = UL/ν, inertia overwhelms viscosity at high Re, producing turbulence.
• Historical detail: Reynolds originally used spark illumination to observe dye clouds, decomposing what looked like a single blob of color into clear curls and vortices.
• The authors also use figures to show the spatiotemporal evolution of turbulence's fine structures being carried, bent, and stretched, and stress that characterizing turbulence requires multiple statistical "rulers": energy spectra, temporal correlations, velocity gradients, and more.

Why it matters

For AI4Science, turbulence simulations are widely used as training and evaluation data sources; the paper shows numerical choices profoundly alter the spatiotemporal coupling dynamics in that data—meaning model training and benchmark evaluations that ignore these choices may be distorted. Treating turbulence as a 4D phenomenon also sets new requirements for model architectures and dataset design.

2026-10-26 ~ 2026-10-26 · 13 related posts

• [source] Johannes Brandstetter opens a turbulence thread: simulated flow quality shapes AI4Science models — jo_brandstetter · 2026-10-26
• Johannes Brandstetter on why numerical choices in turbulence simulation matter for AI4Science training data — jo_brandstetter · 2026-10-26
• Turbulence thread: the convection–viscosity competition inside the Navier–Stokes equations — jo_brandstetter · 2026-10-26
• The contest inside Navier-Stokes: (u·∇)u versus viscosity — jo_brandstetter · 2026-10-26
• Reynolds' spark-lit experiment and the number Re = UL/ν behind turbulence — jo_brandstetter · 2026-10-26
• Watching eddies move, deform and persist as 3D flow evolves — jo_brandstetter · 2026-10-26
• What the "3+1" in (3+1)-dimensional turbulence means — jo_brandstetter · 2026-10-26
• How decades of research characterize turbulence statistically — jo_brandstetter · 2026-10-26
• Kolmogorov 1941: the famous E(k) ∝ k^(-5/3) law of turbulence — jo_brandstetter · 2026-10-26
• Classic spectra plot of eddy correlation times across wavenumbers — jo_brandstetter · 2026-10-26
• [source] Tennekes' 1975 sweeping picture: eddy correlation time scales as 1/(Uk) — jo_brandstetter · 2026-10-26
• Close-up of turbulence: thin structures carried, bent and stretched over time — jo_brandstetter · 2026-10-26
• [source] New paper: 28TB time-resolved DNS shows turbulence must be treated as 4D phenomenon — jo_brandstetter · 2026-10-26