Gaussian process regression for the prediction of aerodynamic performance

Abstract

The development of reusable hypersonic vehicles presents significant challenges due to the complex and computationally intensive nature of high-fidelity simulations required for aerodynamic performance prediction. This thesis explores the use of Gaussian Process Regression (GPR) as a surrogate modelling technique to efficiently and accurately predict the aerodynamic coefficients—namely drag, lift, and moment—of re-entry vehicles such as capsules and gliders. A multi-output GPR architecture is implemented to capture interdependencies between outputs and reduce the number of required simulations. High-fidelity CFD simulations using the DLR TAU code serve as the training dataset for the surrogate models. The study evaluates various kernel functions, sampling strategies, and model configurations to optimize predictive performance, achieving high accuracy with significantly reduced data requirements. Results show that GPR models can reliably predict aerodynamic coefficients across a wide range of flow conditions, with a mean relative error below 1% for drag in realistic re-entry trajectories. This approach enables the rapid generation of aerodynamic databases, offering a valuable tool for early-stage design and trajectory planning of hypersonic vehicles.