Surrogate Modeling of Urban Boundary-Layer Flow

Abstract

Surrogate modeling is a viable solution for applications involving repetitive evaluations of expensive computational fluid dynamics models, such as uncertainty quantification and inverse problems. This study proposes a multi-layer perceptron (MLP) based machine-learning surrogate for canopy flow statistics accommodating any approaching mean-wind angle. The training and testing of the surrogate model is based on results from large-eddy simulations of open-channel flow over and within surface-mounted cubes under neutral ambient stratification. The training dataset comprises flow statistics from various approaching mean-wind angles, and the surrogate is asked to "connect between the dots," i.e., to predict flow statistics for unseen values of the approaching mean-wind angle. The MLP performance is compared against a more traditional spline-based interpolation approach for a range of training data. In terms of relative mean absolute errors on individual flow statistics, the proposed MLP surrogate consistently outperforms the spline interpolation, especially when the number of training samples is reduced. The MLP model accurately captures mean profiles and three-dimensional flow variability, offering robust predictions, even when trained with as few as four approaching wind angles. The model is $10^4 \times$ faster than large-eddy simulations, thus proving effective for multi-query tasks in the context of urban canopy flow modeling.