This paper discusses an emerging area of applying machine learning (ML) methods to augment traditional Computational Fluid Dynamics (CFD) simulations of road vehicle aerodynamics. ML methods have the potential to both reduce the computational effort to predict a new geometry or car condition and to explore a greater number of design parameters with the same computational budget. Similar to traditional CFD meth-ods, there exists a broad range of approaches. In particular, the accuracy and computational efficiency of a CFD simulation vary greatly depending on the choice of turbulence model (DNS, LES, RANS) and the underlying spatial and temporal numerical discretizations. Similarly, the end-user must select the correct ML method depending on the use-case, the available in-put data, and the trade-off between accuracy and computational cost. In this paper, we showcase several case studies using various data-driven ML methods to highlight the promise of these approaches. Whilst these case studies are not comprehensive investigations of the underlying methods and do not include all possible ML approaches (i.e., physics-driven), they highlight the ability of these models to in general predict new designs in near real-time (i.e., less than 5 seconds), after typically less than 1 hour of training on a single GPU. There still exists a need for high quality training data from traditional CFD methods and high-fidelity CFD simulations to validate the ML predictions. Thus, ML approaches should be seen as tools to augment traditional CFD methods rather than to replace them. While this work focuses on preliminary studies, future work will look at more comprehensive real-world/industrial-size calculations for the more promising technologies identified here.
Research areas
