Using Machine Learning Approach to Predict Urban Canopy Flows for Urban Land Surface Modeling

The representation of urban land surface atmosphere interactions is critical in numerical weather prediction and climate models. While recent developments in urban canopy model, the city-specific multi-scale land surface heterogeneities and high computational cost of fine resolution models remain persistent challenges. Using geometry-resolving large eddy simulation (LES) results, we implement a successive encoder-decoder convolutional neural network to develop a computationally efficient trained model to predict the mean streamwise velocity field directly from urban geometric information. The test results show that the trained model can capture the flow patterns well. A suit of systematic evaluation tests for increasingly dissimilar surface geometries shows that success of the trained model can be attributed to prediction of the nonlocal influence induced by tall buildings. The bulk drag coefficient is also evaluated, and the relative error of the trained model is about 32% smaller than the default parameterization in commonly adopted in the single-layer urban canopy model. This proof-of-concept study shows the potential of applying machine-learning based approach to develop city-specific parameterizations.