APPLICATION OF MACHINE LEARNING METHODS FOR AERODYNAMIC DESIGN

In conventional design and optimisation of aerodynamic systems, Computational Fluid Dynamics (CFD) technology has cemented its supporting role as an indispensable resource for analysing different flow field regimes. However, this method involves timeconsuming, memory demanding and computationally costly procedures. In this respect, machine-learning techniques have attracted attention from the engineering community representing a promising alternative for flow field predictions. Over the last decades, numerous architectures have been developed to enhance the algorithm predicting performance to address the limitations encountered in capturing transonic flow characteristics due to its non-linear nature. This research aims to improve an existing deep-learning framework with remarkable capabilities in predicting complex structures of the compressible flow around a civil aero-engine nacelle, such as vortices, boundary layers, wakes and shock waves. The proposed methodology consists of a comprehensive approach that entails analysing the flow aerodynamic, generating the CFD-based dataset, rearranging the efficient domain structure, and training and validating the model for both two- and three-dimensional configurations. Specifically, training results are tested under several conditions by varying the dataset size employed. The accuracy of predictions is eventually assessed according to different criteria presented in other studies, including qualitative flow field comparisons, quantitative evaluation of loss functions and boundary layer profile examinations. The outcomes highlight the prediction capabilities of the deeplearning model, demonstrated by a high level of accuracy achieved for two-dimensional axisymmetric flows and an adequate degree for three-dimensional non-axisymmetric configurations. Furthermore, validating the surrogate model under various training modalities permits the individuation of an optimum data volume range for balancing the input sufficiency and the resulting prediction effectiveness. This study represents a small step forward in adopting these innovative strategies in potential design and optimisation procedures for future compact aero-engine nacelles.