Data-Driven Automotive Aerodynamic Shape Optimization

Aerodynamics plays an important role in the development process of vehicles. At high speeds, aerodynamic drag can make up a substantial percentage of the total driving resistance resulting in reduced efficiency. This reduced efficiency can manifest as increased emissions or a loss of total driving range. Current industry-standard approaches of optimizing vehicle aerodynamics require high-fidelity simulations and wind tunnel tests, which can become expensive when conducting parameter sweeps. We conduct an extensive study of hundreds of commercial automobile geometries and obtain their associated flow fields and aerodynamic performance with high-fidelity large-eddy simulations (LES). Additionally, we demonstrate the use of a machine learning-based approach for analysis and optimization. By leveraging an autoencoder which is augmented with learned estimates of the drag coefficient, we obtain a coherent relationship between global vehicle features and their contribution to aerodynamic characteristics. We consider aerodynamic shape optimization in the identified low-order latent space on an example vehicle to demonstrate that we can leverage such a model. A LES is used to validate the modified geometry obtained from the data-driven optimization.