Surrogate-assisted aerodynamic shape optimization of low-rise buildings with large-eddy simulation validation and flow-mechanism analysis

Abstract

Aerodynamic shape optimization of low-rise buildings requires repeated evaluation of wind-induced forces over many geometric parameters and wind directions, making direct wind tunnel testing or high-fidelity computational fluid dynamics impractical for broad design-space exploration. This study develops a data-efficient surrogate-assisted framework for reducing peak aerodynamic force coefficients of isolated low-rise buildings and investigates the flow mechanisms responsible for the optimized performance. A Kolmogorov–Arnold network is trained using the experiment database to predict the peak drag, cross-wind, and lift-related coefficients as functions of height-to-breadth ratio, depth-to-breadth ratio, roof pitch, and wind direction. The model is compared with backpropagation and radial basis function neural networks and achieves improved accuracy and generalization with substantially fewer trainable parameters. The trained surrogate is coupled with a multi-objective genetic algorithm to identify Pareto-optimal building geometries that simultaneously reduce the three target aerodynamic coefficients over the considered wind-direction range. Representative optimized configurations are then examined using large eddy simulation (LES). The LES results show close agreement with surrogate predictions and reveal that load reduction is associated with modified roof separation, weakened suction zones, and altered wake development around the optimized geometries. The proposed approach links data-driven optimization with flow-physics verification and provides a computationally efficient route for the aerodynamic shape design of isolated low-rise buildings under wind loading.