Research On Large Eddy Simulation Model Of Non-equilibrium Turbulence Based On Machine Learning

Non-equilibrium turbulence is a common flow pattern in daily life and engineering practice,which involves processes such as energy backpropagation and excessive energy dissipation.Its complex flow mechanism has not yet been fully explored.Currently,the main turbulence simulation methods include Direct Numerical Simulation(DNS),Large Eddy Simulation(LES),and Reynolds Average Navier Stokes(RANS).The direct numerical simulation method requires a high resolution computational grid,which consumes enormous computational resources,and cannot meet the computational requirements for complex flow fields at this stage.However,the Reynolds averaging method can only provide the average flow field on a large scale,and cannot simulate the small-scale structure of turbulence,resulting in significant errors in the calculation results.The large eddy simulation method is between the direct numerical simulation method and the Reynolds average method,which can not only reduce the computational complexity but also retain the main turbulence pulsation information.Currently,most traditional turbulence models and subgrid-scale models are constructed based on equilibrium assumptions,which cannot accurately predict non-equilibrium turbulence phenomena,greatly limiting the application of traditional models.Therefore,in order to construct a subgrid-scale model that considers non-equilibrium turbulence,this paper selects the non-equilibrium turbulent flow field as the research object.According to its flow characteristics,the artificial neural network(ANN)algorithm in machine learning methods is used to construct an artificial neural network alternative model for the subgrid-scale model.The main research work of this article is as follows:1.Based on uniform isotropic turbulence data from non-equilibrium turbulence,a subgrid-scale model is constructed using the ANN algorithm to close the subgrid-scale stress tensor.Its input eigenvalue includes the filtering speed,velocity gradient,and the product of the root mean square and strain rate tensor eigenvalues that can characterize the nonequilibrium phenomenon.A priori test shows that the prediction results of the model constructed using neural networks can be highly consistent with the DNS calculation results,and its prediction effect is superior to the traditional gradient model and Samgorinsky model.At the same time,compared to the diagonal component of the subgrid-scale stress tensor,the non diagonal component of the subgrid-scale stress tensor has a greater correlation with the non-equilibrium of turbulence.In addition,in order to further verify the impact of the root mean square product of the strain rate tensor and the strain rate tensor eigenvalue product on the alternative model,this paper also trained a multi-layer ANN model in which the input eigenvalue does not include the root mean square product of the strain rate tensor and the strain rate tensor eigenvalue product,The results show that the performance of the model in capturing a few energy backpropagation phenomena is weaker than that of the ANN model with input eigenvalues containing the root mean square of the strain rate tensor and the product of the strain rate tensor eigenvalues.This indicates that the root mean square of the strain rate tensor and the product of the strain rate tensor eigenvalues are important eigenvalues for training ANN alternative models that consider turbulence non-equilibrium.2.The multi-layer ANN model demonstrates the feasibility of using machine learning methods to construct subgrid-scale models that take into account the non-equilibrium properties of turbulence as an alternative model.However,the multi-layer ANN model has problems such as a large number of neurons,a long training time,and complex model calculations.To address these issues,this paper uses the ANN algorithm to construct multiple single-layer ANN substitution models to further explore the impact of the number of hidden layer neurons and input eigenvalues on the training effect of the ANN substitution model.The prediction results of ANN substitution models with different numbers of neurons in the hidden layer show that the prediction accuracy of ANN substitution models does not improve with the increase of the number of neurons.In this article,ANN substitution models with 500 neurons perform best.The prediction results of neural network models with different input eigenvalues show that when the input eigenvalues are added to the root mean square of the strain rate tensor and the product of the strain rate tensor eigenvalues,the training speed and accuracy of the ANN alternative model are the highest,indicating that the two physical quantities,the root mean square of the strain rate tensor and the product of the strain rate tensor eigenvalues,used to describe the non-equilibrium properties of turbulence,are beneficial for constructing neural network models that consider the non-equilibrium properties of turbulence.