Convolutional neural network-assisted aerodynamic optimization of tall buildings

Wind-induced loading is a critical factor in tall building design, typically governing the structural design. Designers have an opportunity to reduce the wind effects by modifying the external shape of the building through aerodynamic optimization. This study presents an aerodynamic shape optimization (ASO) workflow that reduces the wind-induced forces by modifying the cross-sectional ratio and the corner geometry of a tall building. The workflow utilizes a convolutional neural network (CNN)-based surrogate model to accelerate the optimization algorithm. The CNN-based surrogate model is shown to effectively predict the aerodynamic responses for a wide range of building geometries. The performance of the CNN-assisted ASO workflow is illustrated using two numerical examples. The first example presents a pair of single-objective optimizations to reduce the mean drag and fluctuating lift forces by modification the corner geometry. Reductions of 26% and 51% are achieved for the mean drag and fluctuating lift forces, respectively. The second example is a multi-objective problem which introduced local climatology data to the ASO study to reduce the drag and lift forces by altering the building’s cross-sectional ratio and corner geometries. Both numerical examples illustrate the efficiency of the CNN-based surrogate model with computational speed-ups ranging from 37–61 times compared to traditional optimization workflows.