Data-Driven Reduced Order Modelling for Aerodynamic Shape Optimisation

Despite the increase in computational capabilities, the computational simulation-based approach for aerodynamic shape optimization remains formidable for industrial routine applications. To make this approach acceptable in design practice, data-driven methodologies for building reduced-order models (ROMs) have been proposed. Instead of simplifying the model computed using computational fluid dynamics (CFD), the proposed ROM methodology aims to directly reduce the computational complexity of the model non-intrusively. In this article, we propose a machine learning method for building reduced-order models using multi-variate neural networks and demonstrate how it can maintain accuracy for making predictions in the highly non-linear transonic regime. The pyOptSparse framework is used to provide an interface to the ROM-based optimization process. To demonstrate the convergence, stability, and reliability of the ROM, the ADODG (AIAA Aerodynamic Design Optimization Discussion Group) NACA 0012 first benchmark case of inviscid transonic optimization is extensively studied using the SU2 software package and used for the purpose of validating the proposed ROM methodology.