인공지능 기법과 차수 저감 모델을 이용한 익형 유동장 예측

Abstract

Aerodynamics design optimization with high accuracy and reliability requires huge computational cost because of numerous repetition of CFD. The huge computational cost of CFD based aerodynamics design optimization is the biggest obstacle of practical application. In this research, artificial intelligence techniques and reduced order model are utilized to reduce the computational cost for predicting airfoil flow field. Various airfoil flow fields, including various flow conditions, airfoil shapes, and unsteady vortex shedding, are trained and tested with POD, GPR, CNN, and LSTM methods. As a result, the POD-GPR method showed good accuracy for the dataset with high linearity. On the other hand, the conditional U-net method showed good performance for the dataset with high nonlinearity. In addition, the unsteady flow field is trained and tested with POD-LSTM and conditional U-net method for extrapolation and interpolation of the flow field, respectively. In conclusion, the computational cost for predicting the airfoil flow field is reduced by order of magnitudes with acceptable errors for design optimization.