A systematic dataset generation technique for data-driven automotive drag prediction

Data-driven models for automotive drag suffer from a lack of detailed, freely-available geometries to use as training data, largely due to the proprietary nature of realistic models.

We propose a method for generating a controlled dataset of geometries by barycentric interpolation between a small number of starting cases. Our methods uses ray-tracing to convert from unstructured STL representations of geometry to structured, binary representations. This dataset provides more control over the training process for data-driven drag prediction models, and provides increased granularity that facilitates learning. In addition, the method automates the generation of individual data points, and reduces reliance on existing, openly available data. We test this method on the DrivAer benchmark case, with three different starting configurations - a fastback, a notchback and an estate - to interpolate between. We run the proposed method to generate 1275 data points, and run LES of them using the CharLES solver. Convolution-based networks are then used to make predictions of quantities at three levels: scalar (the drag coefficient), vector (centreline surface pressure), and tensor (full body surface pressure), with excellent results. The method is generalizable to all problems of parametric design optimization.