Application Of Turbulence Modeling Based On Machine Learning

Computational Fluid Dynamics has a wide range applications in engineering,such as aerodynamic,high accuracy of simulation is required,which has a great impact in research of flow and engineering,the designing,analysis and optimization in engineering are driven by accurate results.In this work,the machine learning is used to optimize eddy viscosity models which is shortage of calculating separated flows.The main contents follow: machine learning is used to construct the function between input and output.And the accuracy of regression models is researched under the circumstance that the training set and testing set are obtained from flow fields with different Reynolds number.And the case that training set and testing are obtained from flow fields with different geometry is also investigated.In addition,the approach of bringing back Reynolds stress to field is optimized,more accurate result can be obtained in this way.First of all,the function between input and output is constructed,the accuracy of regression models is researched under the circumstance that the training set and testing set are obtained from flow fields with different Reynolds number.The prediction results of Reynolds stress anisotropy,turbulent kinetic energy and components of Reynolds stress are analyzed.According to the analysis,the following conclusions are drawn:although the discontinuity of results is unsatisfying,excellent predictive performances are observed in Reynolds stress anisotropy,turbulent kinetic energy and components of Reynolds stress.Secondly,the case that training set and testing are obtained from flow fields with different geometry is also investigated.Target flow field is simulated with Reynolds stress transport model.The results of prediction and target flow field are compared,following conclusions are drawn: the results are unsatisfying near the position where a big difference between two geometry.However,excellent predictive performances are observed in the middle of channel,the information of reattach area and boundary layer can be captured by regression models by which the production and dissipation of turbulent kinetic energy and turbulence fluctuation can be predicted.Finally,based on the issue of bringing back Reynolds stress to flow field,the approach is optimized,in this part,the Reynolds stress is separated into linear and nonlinear parts,regression models of Reynolds stress and the linear part are constructed respectively.It turns out that there is a significant improvement of flow field obtained by this way.However,due to the treatment of Reynolds stress,extra regression function need to be constructed,the amount of calculation also increased.