Neural network-based multi-point, multi-objective optimisation for transonic applications

In the context of aircraft applications, the overall design process can be challenging due to the different aerodynamic requirements at several operating conditions and the total associated computational overhead. For this reason, the use of low order models for the optimisation of complex non-linear problems is sometimes used. This paper addresses the challenge of transonic aerodynamic design optimisation through the integration of a set of neural networks for the prediction of integral values, the classification of flow features and the estimation of flow field characteristics. The design method improves the computational efficiency relative to an expensive design process driven by Computational Fluid Dynamics (CFD) evaluations. The approach is used for the multi-point, multi-objective optimisation of a compact aero-engine nacelle in which the design outcomes are validated using a CFD in-the-loop optimisation strategy. It is demonstrated that the method based on the neural network capability identifies similar nacelle designs at a 75% reduction in the overall computational cost, a drag uncertainty prediction within 2.8%, and a predictive accuracy for the classification metric of 98%. For downselected configurations, the main flow characteristics in terms of peak Mach number, pre-shock Mach number and shock location are well predicted by the neural network models compared with the CFD-based evaluations.