What information do Reynolds-averaged models lack to predict smooth-surface separation?
In plain words
Engineering simulations cannot afford to resolve every eddy, so they use approximate models of turbulence. These models still fail to predict reliably where a turbulent flow detaches from a smooth curved surface, which governs stall and drag, and it is not known what information they are missing.
Precise statement
For turbulent boundary layers separating under an adverse pressure gradient from a smooth surface (benchmarks such as $2\mathrm{D}\ \text{and}\ 3\mathrm{D}$ smooth bumps at $\mathrm{Re}_L \sim 10^6\ \text{to}\ 10^7$), identify which non-local or history information absent from single-point Reynolds-averaged equations (for example upstream pressure-gradient history, transport of Reynolds-stress anisotropy, or large-scale unsteadiness) is needed to predict separation and reattachment locations within experimental uncertainty. Answer: the minimal added information, demonstrated by a model that uses it and predicts an unseen geometry blind.
What would settle it
Blind predictions of separation and reattachment on a new benchmark geometry agreeing with experiment within stated uncertainty, by a model whose added inputs are stated.
Status in the literature
Unverified note
Data-driven closures improve agreement on training cases but have not shown reliable transfer to unseen geometries as of 2026.