Towards real-time CFD: novel deep learning architecture for transonic wall-bounded flows

Abstract

Low order models can be used to accelerate engineering design processes. These surrogates should meet the conflicting requirements of large design space coverage, high accuracy and fast evaluation. Within the context of aerospace applications at transonic conditions, this can be challenging due to the associated non-linearity of the flow regime. This work develops a deep learning based method for flow-field prediction. It preserves the spatial resolution of the underlying data, which enables the resolution of the boundary layer. This is an advance relative to current state-of-the-art for transonic flows. The architecture is demonstrated for a complex problem of aero-engine nacelles. The prediction of the primitive flow variables is within a root mean square error of 6×10−5. The model is used to extract the nacelle drag, and its accuracy is about 6.8% relative to CFD computations. The overall method is an enabling and fast preliminary design capability with self-consistent data for multidisciplinary design studies.