An enhanced hybrid deep neural network reduced-order model for transonic buffet flow prediction

Abstract

In the transonic flight environment, transonic buffet has significant influence on the aerodynamic characteristics of aircraft. In the face of the strong nonlinear characteristics of transonic buffet, the reduced-order model (ROM) technology based on deep learning (DL) is introduced to study the flow field. In this article, an enhanced hybrid deep neural network (eHDNN) consisting of Convolutional Neural Network (CNN) and Convolutional Long Short-Term Memory (ConvLSTM) neural network is constructed as the ROM of unsteady transonic buffet flow field. By mining the spatial-temporal characteristic of buffet flow directly from large-scale flow field data, the eHDNN can predict complex flow characteristics such as shock wave motion in the unsteady buffet flow. Structural similarity information and nonlinear improvement strategies are introduced to improve the nonlinear expression ability of eHDNN. The predicted result is consistent with the numerical simulation and has high degree of prediction accuracy. In addition, the eHDNN has good generalization ability, which can predict the spatial-temporal evolution of buffet flow field at unknown angles of attack. The prediction method has good application prospect of the transonic buffet flow control and prediction, which also provides experience for DL modeling in other complex and strongly nonlinear flows.