Comparison of Deep Learning Architectures for Dimensionality Reduction of 3D Flow Fields of a Racing Car

In motorsports, aerodynamic development processes target to achieve gains in performance. This requires a comprehensive understanding of the prevailing aerodynamics and the capability of analysing large quantities of numerical data. However, manual analysis of a significant amount of Computational Fluid Dynamics (CFD) data is time consuming and complex. The motivation is to optimize the aerodynamic analysis workflow with the use of deep learning architectures. In this research, variants of 3D deep learning models (3D-DL) such as Convolutional Autoencoder (CAE) and U-Net frameworks are applied to flow fields obtained from Reynolds Averaged Navier Stokes (RANS) simulations to transform the high-dimensional CFD domain into a low-dimensional embedding. Consequently, model order reduction enables the identification of inherent flow structures represented by the latent space of the models. The resulting data from the 3D-DL study are compared to a traditional dimensionality reduction method, namely Proper Orthogonal Decomposition (POD). Flow field features are examined by using methods of local feature importance, aiming for awareness of predominant fluidic phenomena. We show that our data-driven models capture aerodynamically relevant zones around the racing car. 3D-DL architectures can represent complex nonlinear dependencies in the flow domain. The U-Net network demonstrates an R² reconstruction accuracy of 99.94%, outperforming the results achieved from linear POD with an R² of 99.57%. Efficiently handling numerous CFD simulations leads to improved post-processing and an accelerated investigation procedure for future aerodynamic development. Finally, the discovered findings provide further knowledge for the serial development to increase efficiency, thereby extending, e.g., the range of electric vehicles.