Data driven prediction of roughness-sublayer mean velocity profiles

Most existing roughness-unresolved models of turbulent boundary layers depend on a single roughness length, the equivalent sandgrain height. Though calibration may yield accurate wall shear stress, it does not guarantee accurate prediction of the wall-normal profile of mean velocity below the logarithmic layer, which is important when heat/mass transfer or detailed roughness sublayer (RSL) flow is of interest.

Our earlier work introduced and validated, for a few roughness geometries, an algebraic sublayer velocity profile model (Brereton et al. Phys. Fluids, 33:065121, 2021).

We extend this work to predict RSL velocity as a function of roughness geometrical characteristics, for a wide range of roughness geometries.

Leveraging DNS data of fully developed, fully rough channel flows, deep neural networks, Gaussian process regression, support vector regression are trained to predict the velocity profiles.

These methods are shown to give overall good predictions, though the errors are not small for a few surfaces. Improvement of the predictions may be achieved by integrating additional flow physics. In addition, feature engineering techniques are used to identify roughness features important for the RSL mean flows.