Stop applying AI to every turbulence case: a four-regime map ties local closures to flow physics
Synopsis
AI deployment in turbulence modeling now has a physics-based placement rule. Data-driven Reynolds-averaged closures had been treated as universal mappings and often failed when extrapolated across geometries or forcing conditions, while scale-resolving simulations stay costly. The framework sorts turbulence along two axes, degrees of freedom and departure from equilibrium, and finds that local machine-learning closures transfer only in the near-equilibrium disorganized regime, that strongly coherent flows require explicit scale resolution, and that physics-informed neural networks suit low-dimensional, strongly non-equilibrium deterministic flows.
Interpretation
Whether a machine-learning closure belongs in a given turbulent flow depends on whether that flow admits a coarse-grained statistical description, not on model capacity or data volume. Data-driven Reynolds-averaged closures had been treated as a general mapping from local mean-flow statistics to unresolved Reynolds stresses, and could fail badly when extrapolated across geometries or forcing conditions. The criterion follows from a Mori-Zwanzig projection that splits the unresolved stress into a local Markovian response, a history-dependent memory term from long-lived coherent structures, and residual stochastic forcing, showing the instantaneous local closure to be a restricted form of the full constitutive functional.
Two physical axes place a flow in one of four regimes and thereby indicate which machine-learning approach fits. Classical classification used only geometry and Reynolds number, that is, the number of active degrees of freedom, and left out how far the flow departs from equilibrium. The axes are phase-space dimensionality, parameterized by Reynolds number, and the ratio of mean deformation time to turbulent relaxation time, a turbulent Deborah number; together they form a four-regime matrix covering simplicity, disorganized complexity, organized simplicity and emergent complexity.
In the emergent-complexity regime, where long-lived coherent structures dominate, local closure is mathematically unavailable and the memory-bearing structures should instead be resolved explicitly. Mapping local kinematics to Reynolds stresses is non-injective there, so distinct geometries can share identical local kinematics yet carry entirely different stresses. When training data are aggregated across geometries, the mean-squared-error-optimal predictor collapses to a stress average weighted by how often each local state appears in each geometry, yielding a smeared stress tensor that satisfies no valid governing equations, with a Bayes error that more data diversity cannot remove.
Perspective
The framework addresses closure modeling and machine-learning method selection for incompressible turbulence, written for researchers and engineers in computational fluid dynamics and data-driven modeling. It supplies a qualitative physical partition for judging whether a flow admits a local statistical closure and whether to choose a local machine-learning closure, a physics-informed neural network, or a scale-resolving simulation. Establishing precise quantitative thresholds is stated as the next step, with preliminary work drawing on non-equilibrium thermodynamics toward that goal and toward an adjudicator that evaluates flow states in situ and selects a physically valid, computationally optimal strategy.
The partition rests on qualitative criteria, and specific threshold values of the turbulent Deborah number remain undetermined, which the text lists as future work. The memory kernel is described as an emergent operator tied to the macroscopic flow configuration, so different flows have structurally unique kernels, which limits purely data-driven closures. The text also notes that in out-of-distribution applications neural networks learn statistical correlations rather than invariant causal relationships, so errors can be uncontrolled.
