A data-driven interactive system for aerodynamic and user-centred generative vehicle design

Usama, Muhammad and Arif, Aqib and Haris, Farhan and Khan, Shahroz and Afaq, S. Kamran and Rashid, Shahrukh; (2021) A data-driven interactive system for aerodynamic and user-centred generative vehicle design. In: 2021 International Conference on Artificial Intelligence, ICAI 2021. IEEE, Piscataway, NJ, pp. 119-127. ISBN 9781665432931 (https://doi.org/10.1109/ICAI52203.2021.9445243)

[thumbnail of Usama-etal-ICAI2021-A-data-driven-interactive-system-for-aerodynamic]
Preview
 Text. Filename: Usama_etal_ICAI2021_A_data_driven_interactive_system_for_aerodynamic.pdf
Accepted Author Manuscript
Download (6MB)| Preview

Abstract

In this work, we propose a data-driven design pipeline for quick design exploration of performance and appearance guided alternatives for vehicle design. At the heart of our system is a machine learning-based generative design method to provide users with a set of diverse optimal design alternatives and an interactive design technique to induce users’ preference into the design exploration. The generative design method is the structure on two search process, qualitative and quantitative. To avoid the curse of dimensionality, the qualitative search process first builds up a lower-dimensional representation of a given design space, which is then explored using the unsupervised k-means clustering to synthesise a representative set of user-preferred designs. The quantitative search process explores the design space to find an optimal design in terms of performance criterion such as drag coefficient. To reduce the computational complexity, instead of evaluating drag via Computational Fluid Dynamics simulations, a surrogate model is developed to predict the drag coefficients. The designs generated after the generative design step are presented to the user at the interactive step, where potential regions of the design space are identified around the user-selected designs. Afterwards, a new design space is generated by removing the non-preferred regions, which helps to focus the computational efforts on the exploration of the user preferred regions of the design space for a design tailored to the user’s requirements. We demonstrated the performance of the proposed approach on a two-dimensional side silhouette of a sport-utility vehicle.

CORE (COnnecting REpositories)