An Airfoil Optimization Algorithm Based On Cascade Feedforward Neural Network

In aerodynamic optimization of aircraft configurations,it takes huge amount of computational cost and CPU time for utilizing genetic algorithms.In this work,cascaded feedforward neural network is investigated as a surrogate model for computational fluid dynamics to reduce computational time and to improve computational efficiency.In this work,various methods that can be applied as surrogate models are investigated and compared,including Kriging method,BP neural network,cascade feedforward neural network and so on.With surrogate models constructed by means of these methods,the numerical prediction of flow fields is carried out,while the fitting precision of the corresponding method is compared.Numerical results demonstrate that the cascaded feedforward neural network is feasible for being used as the surrogate model.The shape function of an airfoil is parameterized by the Class-Shape function Transformation(CST)method.For training the cascade network,the point samples of airfoils are randomly generated in a constrained space.The trained network with a required accuracy is then served as the surrogate model to replace the fluid dynamics solver.The lift-drag ratio,calculated by the cascade feedforward network and a fluid dynamics solver,is assigned as the objective function of the single objective genetic algorithm.The CST parameters of the shape function of the airfoil are genetically evolved to optimize the original airfoil.The numerical experiments show that the cascade feedforward network provides the reasonable accuracy with a significant reduction of computing time under a specific objective.