| The flow around the airfoil is relatively simple but widely used in the design of various fluid machinery and aircraft,which is one of the key points of fluid mechanics research.And it can be obtained by the potential flow transformation around the cylinder..Therefore,this paper takes the cylinder and NACA0018 airfoil as the research objects,and used different turbulence models to calculate the flow field of the research object in the steady and unsteady state.The results of numerical simulation is studied through contrast analysis,chooses two-dimensional model SST k-?method calculation results of the research object as the deep learning sample,the flow field of deep learning method retrieval method are studied,selects the multilayer feed forward neural network model to achieve deep learning approach,and based on the flow field calculation of reproduce the airfoil aerodynamic characteristics,main methods and conclusions are as follows:1.The S-A turbulence model,SST k-?turbulence model and LES method are used to simulate the flow field of NACA0018 airfoils at different angl es of attack.The results show that the calculation error of S-A model can be less than 5%at angles of attack in 0-4°.Besides,when the angle of attack exceeds 8°,The S-A and SST k-?numerical simulation methods begin to enlarge errors.When the unsteady method is used,both SST k-?and LES turbulence models can accurately predict the lift coefficient when the angle of attack?is less than 8°.However,due to the flow separation,the SST k-?turbulence model is difficult to predict accura tely when the angle of attack is over 8°.In addition,the LES turbulence model can calculate the drag coefficient of airfoil more accurately.Finally,the LES method around a cylinder is compared with the experimental flow field.By comparing the experimen tal flow field with the numerical flow field,it is found that the SST k-?method could not obtain the process of the vortex around the back pressure surface of the cylinder being pulled and broken by the main stream.It can be concluded that LES has high precision in macroscopic external characteristics and flow field performance by comprehensively comparing the results of three kinds of numerical simulation in the flow around cylinder and airfoil.2.In this paper,the numerical simulation method is use d to obtain the data set,and the quality of the data set directly affects the accuracy of the prediction.The flow around the cylinder is easy to calculate,and the flow around the airfoil can be obtained by potential flow transformation.Therefore,the n umerical simulation method is used firstly to calculate the flow around a cylinder with Reynolds number Re=3900in this paper.By comparing the pressure coefficients of the experimental flow around a cylinder with those of the numerical simulation,it can b e found that the results of LES turbulence model and SST k-?turbulence model are similar,while the calculation time is nearly 35 times different.Considering the calculation time and accuracy comprehensively,the SST k-?turbulence model under the two-dimensional model is selected to obtain the numerical simulation results of the pressure field and velocity field of a cylinder with a diameter of 0.05m under Re=3500-5500.And the results are used as the database of the learning model to study the prediction accuracy of different depths,training steps,learning rates and different network structures.The deep learning results show that when the number of training steps is between 500 and750 steps and the learning rate is 0.1 and remains unchanged,the prediction result is poor.When the learning rate is reduced to 0.01,a better result can be obtained.However,if the learning rate continues to decrease,the neural network is difficult to converge,and it needs to increase the number of training steps t o achieve convergence.In addition,when the number of layers of the neural network increases,the prediction results will be better.However,when the current feed network exceeds 6 layers,the prediction accuracy will not be greatly improved.Therefore,it can be considered that the prediction of flow around a cylinder can be achieved by using a 6-layer feed forward network.Finally,the addition of 0.005 regularization and 1%weight attenuation can effectively prevent overfitting and improve the accuracy of prediction.3.The excellent prediction model of flow around a cylinder is applied to the prediction of flow around an airfoil.Firstly,the numerical simulation method is used to obtain the 140 sets of calculation results of NACA0018 airfoil at angle of attack?=2°,4°,6°,8°,and within the scope of Reynolds number Re=0.1-1.6×106.The calculation results are used as the sample database of deep learning,and the depth prediction model based on feedforward neural n etwork is established.The flow Reynolds number and angle of attack are input into the neural network as independent variable,and the pressure field and velocity field near the airfoil are predicted by neural network calculation.The results show that the calculation results can be obtained by the eight-layer fully connected neural network accurately,and the surface pressure coefficient can be predicted by inverse calculation of lifting resistance coefficient,indicated that the deep learning method can ach ieve good accuracy.By comparing the surface pressure coefficient calculated by XFOIL,numerical simulation and neural network,it is found that the SST k-?turbulence model does not calculate the airfoil separation point,so the neural network could not c alculate the separation point pressure variation,which indicates that the accuracy of deep learning method depends on the database sample.4.By calculating the correlation coefficient between the velocities of the vector field,it can be found that the magnitude correlation coefficient of the velocity vector in the X and Y directions is 0.3,which is relatively weak.Therefore,the independent neural network is adopted to predict the velocity vector field.Compared with the prediction of pressure field and velocity field,because the prediction of the velocity field lacks the constraint between data,the training step length should be increased to more than 600 steps.At the same time,considering that the velocity of the stationary point in the flow field and the velocity in the Y direction of some regions are close to0,the data whose absolute value is less than 10-5 in the calculation results can be returned to zero,and the average error of the vector field prediction can be reduced to less than 7%. |
PDF Full Text Request