Study of shape optimization of an airfoil via machine learning

Abstract

Aerodynamic shape optimization is one of the most extensively studied topic in aeronautical engineering, as an aircraft’s shape is directly linked to its performance. Optimizing the geometry can improve overall efficiency, leading to increased fuel economy, reduced operational costs, and lower emissions. This thesis follows the approach of aerodynamic shape optimization, focusing on enhancing the aerodynamic efficiency of NACA 4-digit series airfoils under different flow scenarios, including Stokes flow and more realistic conditions governed by the Navier–Stokes equations. Unlike traditional adjoint-based gradient optimization methods, this study explores the use of machine learning techniques to estimate the optimization gradient. Specifically, a deep neural network is trained on a custom dataset containing airfoil geometries and their corresponding aerodynamic efficiencies, generated using a self-developed fluid solver under various flow conditions. The numerical framework used in this study is built upon Swan, an open-source software which enables the simulation of various physical problems using the finite element method and allows the integration of certain machine learning techniques, such as the implementation of feedforward neural networks. The results demonstrate that the developed fluid solver performs relatively efficiently for Stokes flow. However, extending simulations to the Navier–Stokes case introduces stability challenges and high computational costs, limiting dataset generation to Stokes flow cases. Consequently, both neural network training and shape optimization are performed exclusively under Stokes flow conditions. The optimized design corresponds to a thin, symmetric airfoil operating at a relatively high angle of attack, consistent with expectations for viscosity-dominated flows.