Title: Machine learning-based multi-objective optimisation of automotive aerodynamic add-ons

Authors: Lingxi Deng; Xueliang Pang; Yanming Du; Xingren Zheng

Addresses: School of Electrical and Information Engineering, Wuhan Institute of Technology, Wuhan, Hubei, 430205, China; School of Mechanical Engineering and Automation, Wuhan Textile University, Wuhan, Hubei, 430200, China ' School of Electrical and Information Engineering, Wuhan Institute of Technology, Wuhan, Hubei, 430205, China; School of Mechanical Engineering and Automation, Wuhan Textile University, Wuhan, Hubei, 430200, China ' School of Electrical and Information Engineering, Wuhan Institute of Technology, Wuhan, Hubei, 430205, China; School of Mechanical Engineering and Automation, Wuhan Textile University, Wuhan, Hubei, 430200, China ' School of Electrical and Information Engineering, Wuhan Institute of Technology, Wuhan, Hubei, 430205, China; School of Mechanical Engineering and Automation, Wuhan Textile University, Wuhan, Hubei, 430200, China

Abstract: This study aims to optimise a vehicle aerodynamic add-on using machine learning and computational fluid dynamics (CFD) to reduce drag (Cd) and lift coefficients (Cl), thereby improving fuel efficiency and stability. Based on the Ahmed model, a duckwing-like add-on was designed with five key geometric dimensions as variables. Among the radial basis, extreme learning machine, support vector machine, and BP neural network models, the back propagation (BP) neural network was found to have the best prediction accuracy. It was then combined with the non-dominated sorting genetic algorithm (NSGA-II) for multi-objective optimisation to generate the Pareto frontier. The results demonstrated significant improvements in aerodynamic performance, achieving up to 24.83% reduction in Cd and 15.5% reduction in Cl. Analysis using extreme randomised tree (ET) and random forest (RF) algorithms indicates that the add-on angle is a critical factor influencing drag coefficient (Cd), while cloud visualisation analysis of the Ahmed model further validated its effectiveness in controlling wake flow. This approach offers a novel method for optimising vehicle aerodynamics through multi-objective optimisation.

Keywords: multi-objective optimisation; machine learning; BP neural networks; genetic algorithms; computational fluid dynamics; CFD; Ahmed vehicle models; aerodynamic add-ons.

DOI: 10.1504/IJVP.2026.155228

International Journal of Vehicle Performance, 2026 Vol.12 No.3, pp.320 - 361

Received: 28 May 2025
Accepted: 24 Sep 2025
Published online: 29 Jul 2026 *

Full-text access for editors Full-text access for subscribers Purchase this article Comment on this article