Parametric Modelling and Variable-Fidelity Bayesian Optimization of Aerodynamics for a Reusable Flight Vehicle

Abstract

A variable-fidelity Bayesian optimization approach that leverages low-fidelity data (panel approach) to efficiently establish an initial prior for aerodynamic optimization of reusable flight vehicles is proposed. This approach demonstrates a notable advantage over traditional Bayesian optimization techniques constrained by their reliance on high-fidelity data and the associated computational expenses. A comparative analysis reveals that our approach can identify the optimized solutions that would typically require a substantial amount of data, using only a limited number of high-fidelity samples. While the traditional approach undergoes significant shifts in the search space over 50 iterations due to Bayesian optimization’s tendency to explore unknown space, our approach, employing low-fidelity data as an initial prior knowledge, achieves stability within approximately 10 iterations. Notably, with just 50 computational fluid dynamics (CFD) samples (high-fidelity data), the optimized vehicle shape demonstrates significant improvements in the lift-to-drag ratio across a broad range of the attack angles, showing a 9% enhancement at the target lift-to-drag ratio at the 10° attack angle, which is the optimization objective.