An Efficient Aerodynamic Optimization Method using a Genetic Algorithm and a Surrogate Model

摘要

A reliable method is presented for robust estimation of the expensive objective functions in single objective optimization algorithm. Multi Layer Perceptron Neural Net (NN) is successfully implemented for evaluating computationally expensive aerodynamic objective functions while the normal distribution concept is applied to determine the parts of the design space which are trained to the NN. Detecting these parts, NN is successfully implemented for evaluating computationally expensive aerodynamic objective functions in design optimization of airfoil shape at viscous transonic flow conditions. This approach, results in more precise NN estimation while decreasing the NN requirements. The accuracy and efficiency of the method is validated with simple Genetic Algorithm. The total number of flow solver calling is noticeably reduced through using this technique, which in turn reduces the total time without deteriorating the optimization algorithm.