Surrogate Based Prediction of 2D Car Wake

In this study, different surrogate models for flow field prediction are applied to a size-limited experimental dataset of a 2D car wake. The dataset consists of DrivAer model wakes influenced by different rear wing positions. The objective is to identify the most robust and accurate method for small training datasets in order to exploit the time-reducing benefits of surrogate models in the experimental field. Three methods (random forest regression, Gaussian process regression and feedforward neural network (FNN) with different activation functions) trained with different dataset sizes are compared. It is shown that Gaussian process regression is the most accurate but most time-consuming method. Interestingly, the random forest regression predicts the flow fields as good as the FNN, but its training time is two orders of magnitude faster. In terms of FNN, rectified linear unit 6 (ReLU6) stood out from the other activation functions as it achieved the best wake predictions with only five training samples.