Research On The Application Of Machine Learning Methods In Surrogate Turbulence Models

RANS model has been widely used because of its ease of use and high efficiency,but the uncertainty of this model will cause the problem of insufficient simulation accuracy.In addition,although LES methods achieve good computational accuracy in high-resolution computations,the large computational resources required limit the use of these methods.Aiming at the problems of traditional turbulence model,this paper uses machine learning method to build machine learning surrogate models for RANS model and LES model respectively,and the main work completed is summarized as follows:1.Comparative study on the construction of surrogate models in RANS turbulence model by different machine learning algorithms.First,starting from the Boussinesq eddy viscosity hypothesis,using a variety of machine learning model and the steady state turbulent eddy viscosity coefficient were used to consturct the surrogate model respectively.A comparative study of model validation is carried out for the classical problem of flows over backward-facing step.The calculation results show that,In terms of computational efficiency,the models constructed by each machine learning algorithm can be effectively improved.In terms of calculation accuracy,the calculation of physical quantity of each algorithm surrogate model can be guaranteed within the ideal range,and the machine learning surrogate model can effectively capture the phenomenon of separation and reattachment.Among them,artificial neural network,extreme gradient boosting and K-nearest neighbor algorithm have more potential in constructing machine learning surrogate RANS turbulence model.2.Research on the generalization of surrogate machine learning models under different Reynolds numbers.The simulation results in the test flow with different Reynolds number distribution ranges relative to the training flow show that when the test flow Reynolds number is close to the training flow Reynolds number,the simulation results of the surrogate model can effectively match the original RANS simulation results.When the Reynolds number of test flows is lower than the Reynolds number of training flow,there will be a deviation,and this deviation becomes more obvious as the Reynolds number decreases.In the test flow with a higher Reynolds number relative to the training flow,the simulation deviation will be further reduced as the Reynolds number increases.It is revealed that the computational accuracy of the machine learning surrogate RANS model increases with the increase of Reynolds number in the test flow.3.For large eddy simulation,machine learning methods are used to reconstruct the research of the sub-grid scale unclosed term.The model validation studies were carried out in channel flows at different Reynolds numbers.Through the study of numerical simulation of turbulent channel flow at a low Reynolds number,it is shown that compared with the calculation of the original dynamic Smagorinsky model,the new machine learning sub-grid scale surrogate model can accurately predict the statistical properties and spatial structure of the turbulent flow field,and is more efficient in calculation.There is a significant improvement.It shows that the application of machine learning algorithms in large eddy simulation optimization modeling has high significance,and has broad application prospects for revealing complex turbulent flow phenomena.