Machine learning modeling for the near-wake mean velocity deficit profiles behind a rough circular cylinder
Junyi Mi,Shujin Laima,Xiaowei Jin,Hui Li
DOI: https://doi.org/10.1016/j.oceaneng.2022.111791
IF: 5
2022-09-01
Ocean Engineering
Abstract:This study proposes a machine-learning framework to model the near-wake mean velocity profile, illustrating the similarity in the near wake of a circular cylinder. A dataset comprising the near-wake velocity field and wall pressure of the rough circular cylinder, ranging from the subcritical to supercritical flow regimes, is obtained via the wind tunnel test. Then, singular value decomposition of the mean velocity deficit matrixes is performed to obtain the left- and right-singular vectors. The generalized basis and coordinate are finalized by analyzing the scatter of the singular vectors, the correlation of the singular vectors with inputs, and the energy distribution of different generalized basis and coordinates. Next, the random forests (RF) algorithm is applied with the asymmetry of the wall pressure coefficient, drag force coefficient, Reynolds number, and spatial coordinate as inputs, and outputs the generalized coordinates. Finally, the importance of inputs is investigated through Shapley Additive Explanation (SHAP), which enhances the interpretability of the model. The test results show that the RF model can accurately predict the near-wake mean velocity deficit profiles at various Reynolds numbers. The SHAP evaluation indicates that the asymmetry of the pressure coefficient is crucial for modeling the velocity fields among different regimes.
engineering, civil, ocean, marine,oceanography