Research On Flow Characteristics Prediction And Optimization Of Hydrofoil Based On Deep Learning

Hydrofoil,an extremely important baseline element for the power or energy-saving devices in the field of hydraulic machinery,has a wide range of engineering applications,e.g.,water turbine,water pump,marine propeller,pump jet propeller,etc..Preventing from the adverse effect,including the hydrodynamic performance degradation,stress deformation and fatigue damage of hydrofoils,caused by the geometric design failure is a tough problem that must be addressed for its real engineering applications.Hence,the investigation on how to maximize the operation efficiency of the fluid machinery and to solve the problem of the operational instability and fatigue damage caused by the cavitation erosion and stress deformation is the primary objective of the hydrofoil design.In this regard,the research on the rapid and accurate prediction of the hydrodynamic characteristics of hydrofoils can provide an essential and helpful reference for its structural design,which leads to the enhanced geometrical optimization efficiency and hence is of great significance to improve the hydrodynamic performance of the devices.With the breakthrough development of the machine learning in various research fields,a prediction method based on the machine learning algorithm has gradually entered people’s visions.This new method absorbs the complex geometric structure and hydrodynamic performance of hydrofoils as the learning object,establishing the prediction model through neural network to realize the prediction of unknown hydrofoil performance.Its main advantages lie in: 1)can greatly reduce the calculation cost and prediction time;2)can maintain the insights into the characteristics of any unknown hydrofoils through the fast flow field prediction;3)can realize the rapid adjustment of hydrofoils while maintaining the prediction ability with the same accuracy as the traditional tools of the numerical simulation and experimental test.In this thesis,the general physical phenomena of airfoil flow is investigated by adopting the deep learning methods.As a result,the rapid and accurate prediction of the flow dynamic performance of hydrofoils is achieved and naturally based upon which,the geometrical optimization design of the hydrofoil is carried out,with some innovative results obtained.The main achievements of this thesis include:1.Establishment of airfoil flow database.Firstly,the automatic airfoil calculation platform is established by Python script,and through the computational fluid dynamics(CFD),the airfoil coordinate point database of University of Illinois at Urbana Champaign(UIUC)is simulated.Specifically,by applying the k-ω SST turbulence model in the open source CFD software Open FOAM,the flow characteristics around different types of 1550 airfoils under 26 different attack angles are obtained.With the deep learning training samples generated,the flow data around different airfoils are analyzed.The results show that most of the airfoils in the UIUC database exhibit the flow separation phenomenon after the attack angle is beyond 11°.From which,the difference of lift coefficient for each airfoil begins to increase,while the drag coefficient increases more prominently.With the further increase of the attack angle,the varying level of flow stall phenomenon occurs for all selected airfoils in the database.2.CNN prediction and optimization of airfoil hydrodynamic parameters.Based on the convolutional neural network(CNN),three prediction models are proposed to predict the hydrodynamic parameters of airfoils,including the lift coefficient,drag coefficient and pressure coefficient.The data set includes 1550 airfoils in UIUC at the attack angle of-5°,-4°,…,20° and the inlet velocity of 1 m/s.Taking the geometric characteristics and hydrodynamic parameters of the airfoil as the learning objects,the procedures of CNN modeling,training and prediction are carried out.The results show that the CNN prediction model has a high computational efficiency for predicting the hydrodynamic parameters of the airfoil with good accuracy.Based on the prediction model,the lift coefficient and drag coefficient of airfoils are optimized.After optimization,50 groups of Pareto frontier points are obtained followed by the performance analysis of each optimized airfoil.It is reported that the optimization efficiency of the new CNN method is at least one order of magnitude higher than the traditional CFD simulation.3.CNN-FEM fluid structure prediction and structural stress optimization of airfoils.On top of the CNN prediction model of the pressure coefficient on airfoil surface and the finite element method(FEM),the unidirectional fluid structure interaction calculation of airfoil at the fixed angle of 10° is carried out,by comparison to the traditional unidirectional fluid structure interaction calculation results based on CFD and FEM.Results show that the standard error of equivalent force distribution predictions is less than 5%.Based on the new CNN-FEM method,the multi-objective optimization of the airfoil geometry is conducted with the lift drag ratio and the average equivalent stress of the two-dimensional airfoil taken as the optimization objectives for the first time.As a result,the optimized airfoil shape is achieved with a high lift-drag coefficient and high structural strength.4.CNN flow field prediction around airfoils.Based on the U-net network structure,the flow field characteristics of airfoils are predicted.The velocity field and pressure field of airfoils under the fixed attack angle of 10° are predicted with 200 sets of test data used to verify the model.The results show that the prediction model can rapidly establish the mapping relationship between the airfoil structure and flow field information,while the prediction error is extremely low with mostly less than 2%.Within some particular regions such as in the wake and around the wall,the local error is approximately 8%,which is still fairly low proving its capability to accurately predict the flow field characteristics of airfoils.To sum up,the deep learning method provides as an accurate and efficient means of predicting the airfoil flow characteristics.On top of it,the incorporation of the deep learning method into the airfoil shape optimization boosts the optimization efficiency to a large extent,facilitating to achieve a more effective(probably the optimum)airfoil design.Moreover,the realization of the rapid and accurate prediction of airfoil flow field can greatly speed up the evaluation process,which in the future might turn around the long and costly R&D cycle through traditional CFD numerical simulations or experiments,and lay a research foundation for solving any problems related to the airfoil flow characteristics,cavitation,fluid solid coupling,etc..