Deep-learning methods for non-linear transonic flow-field prediction

Abstract

It is envisaged that the next generation of ultra-high bypass ratio engines will use compact aero-engine nacelles. The design and optimisation process of these new configurations have been typically driven by numerical simulations, which can have a large computational cost. Few studies have considered the nacelle design process with low order models. Typically these low order methods are based on regression functions to predict the nacelle drag characteristics. However, it is also useful to develop methods for flow-field prediction that can be used at the preliminary design stages. This paper investigates an approach for the rapid assessment of transonic flow-fields based on convolutional neural networks (CNN) for 2D axisymmetric aeroengine nacelles. The process is coupled with a Sobel filter for edge detection to enhance the accuracy in the prediction of the shock wave location. Relative to a baseline CNN built with guidelines from the open literature, the proposed method has a 75% reduction in the mean square error for Mach number prediction. Overall, the presented method enables the fast prediction of the flow characteristics around civil aero-engine nacelles.