Transonic Airfoil Shape Optimization Using Deep Neural Networks and Genetic Algorithms

This paper summarizes a surrogate-assisted framework for optimizing a transonic RAE۲۸۲۲ airfoil with a deep neural network (DNN) coupled to a genetic algorithm (GA). Five Hicks–Henne bump amplitudes are used to modify the upper surface, while Mach number and angle of attack are treated as operating variables. A Taguchi orthogonal design is used to construct a compact dataset of ۱۲۵ inviscid OpenFOAM/rhoCentralFoam simulations. The resulting seven-input DNN predicts lift and drag coefficients and replaces repeated CFD evaluations inside the optimization loop. The model reaches mean absolute errors of ۰.۵۱۲۲% for lift and ۵.۱۷۲% for drag, while reducing the reported evaluation time from about ۲۰۰۰ s for CFD to ۰.۰۱۳ s per surrogate prediction. Across Mach ۰.۷۰–۰.۸۰ and angles of attack of ۱.۵°–۲.۵°, the best optimized case increases lift by ۲.۹۹%, reduces drag by about ۴۰%, and raises aerodynamic efficiency (Cl/Cd) by ۷۱.۵۴% relative to the baseline. The improvements are associated with weaker and downstream-shifted shocks and smoother pressure variation over the upper surface. The conference version intentionally condenses the full methodology and literature review while retaining the principal computational evidence and optimization findings.

کلیدواژه ها:

Transonic flow ، Hicks–Henne parameterization ، Taguchi DOE ، deep neural network ، genetic algorithm ، surrogate optimization