Evolutionary Multiobjective Aerodynamic Design Optimization Using CFD Incorporating Deep Neural Network

Abstract

We propose to use computational fluid dynamics (CFD) incorporating a deep neural network (DNN) to reduce the computational time of evolutionary multi-objective aerodynamic design optimization. In the proposed approach, DNN infers the steady-state flow field of each design candidate, and the CFD is started from this inferred flow field close to the steady state to reduce the required computational time of CFD. In this paper, this method is applied to a multi-objective airfoil design optimization problem using a multi-objective evolutionary algorithm. The result shows that the total computational time was reduced by 42.0% under a 1-core CPU condition and 13.5% under a 96-core CPU condition by the proposed method.