Multi-fidelity bayesian optimization of unsteady cavity flow

Abstract

This dissertation presents a computationally efficient optimization methodology for reducing aeroacoustic noise generated by unsteady cavity flows representative of internal weapon-bay configurations operating under transonic conditions. The cavity flow is treated as a time-dependent, compressible, and turbulence-dominated problem, where accurate prediction of both tonal and broadband noise components requires advanced numerical modeling techniques. A key challenge in high-accuracy aeroacoustic optimization is the prohibitive computational cost associated with fully unsteady three-dimensional simulations. To address this challenge, a multi-fidelity Bayesian optimization framework combining numerical models of different accuracy levels is developed. Lower-cost simulations are employed to enable efficient exploration of the design space, while higher-accuracy simulations are selectively used to ensure physically reliable evaluation of promising designs. This strategy allows effective optimization to be performed under a fixed computational budget. In addition, the optimization strategy is implemented using both single-fidelity (SF) and multi-fidelity (MF) Bayesian optimization frameworks, and their performance is systematically compared. In both approaches, Gaussian Process-based surrogate models are employed to represent the underlying design response and guide adaptive sampling. In the SF framework, the optimization process is performed using only LF simulations, while HF analyses are employed solely for validation of the optimal designs. In contrast, the MF framework incorporates both LF and HF models within the optimization loop, enabling a more accurate and efficient exploration of the design space. By exploiting the correlation between fidelity levels and incorporating cost-aware acquisition strategies, the MF approach enables faster convergence under a fixed computational budget. Comparative results show that the MF framework achieves improved design quality with fewer HF evaluations, demonstrating a clear computational advantage over the SF strategy for high-cost unsteady aeroacoustic simulations. The proposed methodology is validated using a generic cavity geometry for which experimental reference data are available. Optimization results demonstrate that the framework achieves significant reductions in cavity-generated noise across a wide frequency range, including the dominant resonance components. Analysis of the optimized configurations shows that the achieved noise reduction is associated with modifications to the shear-layer development and a weakening of the acoustic feedback mechanism at the cavity aft wall. In addition to improving aeroacoustic performance, the study highlights the critical role of numerical discretization and mesh quality in unsteady noise prediction. An automated, parametric mesh-generation approach is introduced to ensure consistent and repeatable simulations throughout the optimization process. This capability is essential for maintaining numerical stability and reliability when evaluating a large number of geometrically varying designs. Beyond the specific cavity application, the methodology developed in this dissertation constitutes a general and transferable MF Bayesian optimization framework. The proposed approach can be readily extended to other unsteady flow problems in aerospace engineering, including applications involving aerodynamics, aeroelasticity, structural considerations, and flight-performance analysis. As such, the work provides a scalable and practical design strategy for the optimization of complex aerospace systems subject to stringent computational constraints.