This paper asks whether a machine can learn the probability law that governs turbulent channel flow, and whether samples drawn from that learned law look and behave like real turbulence. The authors train a generative machine-learning model to approximate the probability distribution of flow states for a channel flow at friction Reynolds number Reτ = 180. They then test whether the learned distribution gives the right ensemble statistics, whether it can be sampled consistently when conditioned on partial observations, and whether it respects the flow’s dynamics.

To keep the learning problem manageable, the team trains the model on what they call a minimal conditional flow unit. This is the smallest spatial domain for which the flow outside that box becomes effectively independent of a single observation at the box centre. In other words, outside that domain conditional flow fields look like unconditional ones in a mean-square sense. The authors use a flow-based generative model trained on data from that unit and devise a procedure to draw consistent conditional samples from the learned distribution.

At a high level, the model learns a mapping from simple random inputs to realistic velocity fields. The paper emphasizes conditional sampling: given a local observation, the model can produce many full-field realizations that are all consistent with that observation and with the learned statistics. This capability is important for reconstruction problems, where only sparse measurements are available, and for generating synthetic inflow or initial conditions for simulations.

When the authors compare samples from their model with direct numerical simulation (DNS) data, they find encouraging agreement in several ways. The synthetic fields reproduce key statistical and dynamical signatures of turbulence reported in the paper, including intermittency (irregular bursts of intense activity) and nonlinear energy transfer between scales. They demonstrate the conditional sampling in a flow reconstruction task and use the model to fill a larger domain with synthetic turbulent velocity fields. Importantly, when those synthetic fields are used as initial conditions in DNS, the simulations evolve to physically plausible and statistically stationary ensembles, which suggests the learned distribution is close to the flow’s natural distribution.