深層学習による後流積分領域推定手法の提案

抄録

In the field of computational fluid dynamics, wake integration method has the advantage that can reduce non- physical drag effect by numerical viscosity. However, in order to define its appropriate integration region, visualization of physical quantities and quantitative evaluation are required. In this study, we propose CNN (Convolutional Neural Network) model to simplify the definition of the appropriate region. Data sets were created by computing 2D fluid analysis on 15 types of NACA airfoil and learning data were classified into three groups to compare the airfoil shapes contributing to generalization performance. The drag of the surface integral method was used as the training data and the entropy drag visualization image was used as input to learn and infer with the CNN model, and the drag distribution was predicted. As a result, the prediction error of the drag distribution was 1.19 (%) and it was clarified that the presence of a symmetric wing contributes to the generalization performance of the model.