Gaussian Process Surrogate Model for eVTOL Propeller Aerodynamics

Gaussian Process Regression (GPR) is a flexible, non-parametric machine learning method well-suited for regression tasks. In the context of modeling aerodynamic propellers, GPR significantly reduces the amount of computationally expensive training data needed compared to simpler interpolation or curve-fitting approaches for the same level of accuracy. This work explores several strategies for building a surrogate model of an isolated propeller for the Joby Aviation tilt-propeller electric vertical take-off and landing (eVTOL) aircraft. To better capture sharp local variations in output quantities of interest and accommodate unevenly spaced training data, a novel delta-layer GPR approach is introduced. This method builds on the traditional single-layer GPR method by fitting to the error between the training data and the first layer fit. In parallel, a multi-fidelity GPR model is developed, using lower-fidelity data to achieve better prediction of the underlying mean function while incorporating high-fidelity CFD data for precision. This approach is further extended by integrating a third source of high-fidelity wind tunnel data, resulting in a smooth and accurate surrogate model across the entire flight envelope.