Surrogate Modeling of Urban Boundary-Layer Flow

Surrogate modeling is a valuable tool for applications that require repetitive evaluations of expensive simulations. This work compares two machine-learning surrogates to predict canopy flow statistics, accommodating any approaching mean wind angle. The first surrogate is based on the K-nearest neighbors, and the second on a multi-layer perceptron network. The surrogates are trained and evaluated using flow statistics from large-eddy simulations of open-channel flow over and within an array of surface-mounted cuboids spanning a range of approaching wind angles. The proposed surrogate models are highly efficient and able to reconstruct time-averaged three-dimensional flow statistics with a coefficient of determination > 0.96 when trained using many training samples. As the number of training samples decreases, the accuracy of both models deteriorates, with the MLP featuring an overall superior performance. It will be shown that the observed behavior can be attributed to the non-linearity of the MLP's activation function, which allows it to effectively capture non-linearities in the input-output mapping. Further, this study found that ML-specific metrics may lead to misplaced confidence in model accuracy.