An active learning approach for the prediction of hydrodynamic roughness properties

Realistic surfaces of flow-related equipment are often hydraulically rough due to wear or fouling. Predicting the skin friction exerted by such rough surfaces is a challenging task since the topography of these surfaces is inherently irregular and complex. Recent developments in data-driven methods and increasing affordability of high-fidelity direct numerical simulations (DNS) have created new possibilities for estimation of drag on irregular rough surfaces. In the present work we aim to demonstrate a viable approach to efficiently train a predictive model for the estimation of drag for irregular roughness based on its height probability density function (PDF) and the surface height power spectrum (PS). An active learning (AL) framework is employed to efficiently navigate the construction of a training database. Training data is generated by conducting direct numerical simulations of a flow over artificially generated rough surfaces in minimal channels in order to minimize the computational effort. An ensemble neural network (ENN) model is trained based on the database. The ENN model shows promising potential in predicting the skin friction as well as estimating the epistemic (model) uncertainty. ... mehrFurthermore, the model – trained on artificial surfaces – is tested on five realistic surface scans, showing that a maximum error of 8.7% between the predicted roughness function ∆U+ and the DNS results is achieved. Overall, the AL framework shows a great potential as a basis for future research towards a universal predictive tool for any arbitrary roughness.