Aeroacoustics and aerodynamics optimization with using machine learning algorithms

Abstract

This dissertation aims to develop an innovative optimization methodology by integrating machine learning algorithms with aeroacoustic optimization. To achieve this goal, a general optimization framework has been established, incorporating a Generative Adversarial Network (GAN) algorithm for dimensionality reduction and integrating computational tools for performance evaluation. Through this approach, a high-performance optimization tool has been developed, effectively combining aeroacoustic optimization with GAN-based methodologies. The effectiveness and efficiency of the proposed optimization methodology have been evaluated through various analyses. To this end, four different optimization tools have been developed based on the same methodological framework but incorporating different performance evaluation methods. Comparative analyses of these tools have revealed the impact of computational accuracy on the optimization process, demonstrating that higher-fidelity tools significantly enhance optimization success and efficiency. To investigate the applicability of the GAN algorithm for aeroacoustic optimization, an inviscid-based Noise-GAN optimization tool was developed, utilizing low-fidelity aerodynamic solvers without viscous effects. Due to its fast computational capabilities and reduced complexity, the tool was rapidly integrated into the methodology and tested for its potential in aeroacoustic optimization. Ten independent optimization studies were conducted under specific flight conditions, and the optimized airfoils were compared against profiles from the UIUC airfoil database. The results demonstrated that the Noise-GAN-generated profiles exhibited superior hydrodynamic and hydroacoustic performance compared to existing designs. To improve performance prediction accuracy, a viscous-based Noise-GAN optimization tool was developed. By replacing low-fidelity solvers with high-fidelity viscous solvers, this tool overcame the limitations of the inviscid-based approach. Comparative studies showed that neglecting viscous effects could misguide the optimization process, reducing efficiency and overall performance. A detailed analysis across multiple angles of attack (AoA) revealed that viscous-based Noise-GAN consistently produced high-performance profiles, whereas the inviscid-based method, while capable of generating high-performing designs, exhibited lower efficiency. To extend the methodology to three-dimensional geometries, a 3D Viscous-Based Noise-GAN optimization tool was developed. This tool facilitated the optimization of a hydrofoil at multiple AoA conditions, yielding significantly improved aerodynamic and aeroacoustic performance compared to the baseline NACA0009 profile. To assess the operational limits of the developed methodology, an optimization study was conducted on a helicopter rotor in hover under transonic flight conditions. The rotor was divided into five sections for independent 2D optimizations, and the optimized airfoils were used to construct a 3D rotor geometry. Comparative analyses with a reference rotor showed that the optimized design generated twice the thrust while maintaining a higher thrust-to-torque ratio (CT/CQ). Additionally, the new rotor exhibited a significant noise reduction of up to 14 dB in near-field observer locations, confirming its superior aeroacoustic performance. Finally, the DynStall-GAN optimization tool was developed to enhance wind turbine airfoil performance by delaying dynamic stall. Ten independent optimization cycles were performed under two different conditions, and the resulting profiles outperformed both the NACA0012 airfoil and a previously optimized profile from the literature. A comprehensive evaluation of the conducted studies confirms that the developed optimization methodology is highly effective and versatile. Its compact and adaptable structure has enabled the creation of multiple optimization tools and facilitated the investigation of various aerodynamic phenomena. With the successful implementation of five optimization studies—ranging from 2D hydrofoil optimization to 3D transonic rotor optimization—this research has yielded numerous high-performance airfoil geometries, contributing significantly to the literature on aerodynamic and aeroacoustic optimization.