Machine-learning wall-model large-eddy simulation accounting for roughness

Wall-modeled large-eddy simulation (WMLES) has emerged as a computationally effective framework for modeling the impact of roughness on outer flow without the need to resolve the small-scale flow and roughness geometry near the wall. Our objective here is to develop a robust rough-wall model for WMLES using machine learning (ML) techniques, capable of accurately handling a wide range of flow conditions, including both attached and separated flows, as well as various rough surfaces spanning the transitionally and fully rough regimes. To this end, we compiled an extensive DNS roughness database that encompasses irregular rough surfaces with different distributions of probability density functions and power spectra at various Reynolds numbers. The database serves as the foundation for training our ML-based wall models. The choice of non-dimensional input features for the wall model including both flow variables and roughness parameters is optimized to enhance model performance. The performance of the model is evaluated a-posteriori in WMLES of turbulent channel flows with rough walls. The results demonstrate that our wall model is able to accurately predict drag for both unseen fully and transitionally rough cases. The predictive capabilities of the wall model are also evaluated in a real flow scenario involving a geometrically complex high-pressure turbine blade with roughness.