Aerodynamics-guided machine learning for design optimization of electric vehicles

Area of Science:

• Automotive Engineering
• Computational Fluid Dynamics
• Machine Learning

Background:

• The shift to electric vehicles (EVs) necessitates new design approaches due to the absence of combustion engines.
• Optimizing vehicle aerodynamics is crucial for EV efficiency but computationally expensive with current methods.
• There is a need for efficient surrogate models to aid aerodynamic analysis and design.

Purpose of the Study:

• To develop a data-driven surrogate model for predicting aerodynamic performance of automobile geometries.
• To leverage machine learning for efficient aerodynamic design optimization of electric vehicles.
• To reduce the computational cost associated with traditional aerodynamic simulations.

Main Methods:

• Analysis of a dataset of industry-quality automobile geometries and their aerodynamic performance from large-eddy simulations.
• Application of a nonlinear autoencoder trained to extract low-dimensional representations of geometry and predict drag coefficient.
• Aerodynamic design optimization in the learned low-order latent space.

Main Results:

• A nonlinear autoencoder successfully extracted a low-dimensional relationship between vehicle geometry and aerodynamic performance.
• The model accurately estimated drag coefficients from latent variables.
• Optimization in the latent space yielded designs with improved aerodynamic trends, validated by simulations.

Conclusions:

• Data-driven approaches, specifically nonlinear autoencoders, can effectively analyze and optimize vehicle aerodynamics.
• This method offers a computationally efficient alternative to traditional simulation-heavy workflows in automotive design.
• The findings support the application of machine learning in production environments for vehicle design and analysis.