Deep-learning for flow-field prediction of 3D non-axisymmetric aero-engine nacelles

Computational fluid dynamics (CFD) methods have been widely used for the design and optimisation of complex non-linear systems. Within this context, the overall process can typically have a large computational overhead. For preliminary design studies, it is important to establish design capabilities that meet the usually conflicting requirements of rapid evaluations and accuracy. Of particular interest is the aerodynamic design of components or subsystems within the transonic range. This can pose notable challenges due to the non-linearity of this flow regime. There is a need to develop low order models for future civil aero-engine nacelle applications. The aerodynamics of compact nacelles can be sensitive to changes in geometry and operating conditions. For example, within the cruise segment different flow-field characteristics may be encountered such as shock-wave boundary layer interaction or shock induced separation. As such, an important step in the successful design of these new architectures is to develop methods for fast and accurate flow-field prediction. This work studies two different metamodelling approaches for flow-field prediction of 3D non-axisymmetric nacelles. Firstly, a reduced order model based on an artificial neural network (ANN) is considered. Secondly, a low order model that combines singular value decomposition and an artificial neural network (SVD+ANN) is investigated. Across a wide geometric design space, the ANN and SVD+ANN methods have an overall uncertainty in the isentropic Mach number prediction of about 0.02. However, the ANN approach has better capabilities to predict pre-shock Mach numbers and shock-wave locations.