Multi-objective optimization of high Mach waverider based on small-sample surrogate model

Abstract

Advancements have been achieved in the optimization of waverider designs with the aid of machine learning to expedite the design process. However, these approaches are hampered by the need for extensive sample sizes and susceptibility to becoming ensnared in local optima. This study undertakes a parametric design based on the wedge-derived, power-law-shaped waverider, increasing configuration diversity and creating a dataset with limited samples by calculating waverider geometry and aerodynamic parameters. At a Mach number of 10, a multi-objective optimization design is implemented using the Young's double-slit experiment-least squares support vector regression (YDSE-LSSVR) surrogate model in conjunction with improved congestion distance multi-objective particle swarm optimization algorithm, focusing on maximizing the lift-to-drag ratio and volumetric efficiency as much as possible. The results indicated that, under conditions of limited samples, the YDSE-LSSVR model outperforms standard models such as support vector regression, LSSVR, Kriging, and Polynomial Chaos Expansions-Kriging regarding prediction accuracy. The Pareto solutions for both concave and convex waveriders, obtained through multi-objective optimization, improve the lift-to-drag ratio by 17.36% and 21.70%, respectively, and increase the volumetric efficiency by 88.89% and 105.56%, in comparison to baseline configurations. In addition, the research examines the impact of various design parameters on the Pareto solutions. Finally, the study applies the K-means method to conduct a cluster analysis of the Pareto solutions, generating three-dimensional waverider configurations based on distinguished solutions from different clusters.