Multi-fidelity Gaussian process surrogate models for numerical acoustic design in frequency domain

This thesis investigates Multi-Fidelity (MF) Gaussian Processes (GP) as surrogates in numerical acoustic design. Here, the parameter dependence of the acoustic target quantity is learned based on data of varying fidelity using GP regression. Data of different fidelity levels is generated using finite element models of varying numerical resolution. The fundamental suitability of these data for MF models is examined before discussing the theoretical potential of the method and various practical implementation options. The results of the study demonstrate that while MF models can outperform conventional approaches significantly, their application is considerably more complex and error-prone.