Development of a method for deriving vehicle shapes and flow fields for low aerodynamic drag using machine learning

Abstract

Due to rapid changes in the market and various customer values, it is necessary to shorten the development period of automobiles. Vehicle performance has been improved mainly through experimental analysis and CAE (computer aided engineering). In the increasingly rapid development of vehicles, machine learning is taking the lead. Vehicle performance prediction usually involves constructing surrogate models using design parameters and CAE results. However, three-dimensionally complex vehicle shapes cannot be fully represented by design parameters. Additionally, reducing vehicle drag has become even more important with the rise of battery electric vehicles. While balancing design and vehicle performance, complex three-dimensional shapes are explored to find optimal solutions, which requires a substantial amount of effort. For predicting vehicle drag performance, a method using variational autoencoder (VAE) is proposed to predict the shape of the vehicle front bumper side, drag coefficient, and flow field on the side and rear of the vehicle. With the proposed method, the drag coefficient was predicted with a maximum error of 0.012, an average error of 0.002, and an R² value of 0.88, demonstrating good agreement with CFD (computational fluid dynamics). Additionally, the predicted velocity magnitude distribution on the side and rear of the vehicle is similar to CFD results. By creating a scatter plot (map) of the latent variables of the proposed method using principal component analysis results, it was found that the direction of increase in the first and second principal components corresponded with the increasing trend of the drag coefficient. Using this map, it becomes possible to predict the drag coefficient, velocity magnitude distribution on the side and rear of the vehicle, and vehicle shape derived from the proposed method. By using the proposed method, it is possible to suggest three-dimensional shapes that were not represented by traditional design parameters, making it easier to balance design and performance and thereby facilitating the search for optimal solutions.