Study of Data-Driven Prediction of Roughness Skin Friction

The potential of developing a machine learning predictive model for the roughness equivalent sand-grain size $$k_s$$ k s is assessed in the present work. The training dataset for the machine learning model is obtained by carrying out Direct Numerical Simulations (DNS) on the artificially generated roughness at friction Reynolds number Re $$_\tau =800$$ τ = 800 . The generation of the roughness is based on a mathematical algorithm in which the roughness power spectrum (PS) as well as the height probability density function (PDF) can be prescribed. A roughness repository that contains 4000 artificial roughness is constructed with the generation algorithm. 50 roughness out of the repository are used for training the model. The selection of the training roughness samples is made following the active learning (AL) framework, where the AL framework selects roughness samples based on the model uncertainty. The ensemble neural network (ENN) model is employed to quantify the model uncertainty. Excellent performance of the ENN model is observed in the present work. Eventually an averaged prediction error of 8.5% is achieved for the ENN model tested on 4 realistic surfaces.