Surrogate modelling of unsteady wind pressure on tall buildings with varying slenderness and side ratios using attention-based neural networks

Machine learning techniques have become increasingly prevalent in wind engineering applications for tall buildings. However, most existing studies have focused on predicting either static pressure or time-dependent pressure on a single tall building. In this study, a machine learning-based surrogate model was developed to predict wind pressure time histories on rectangular tall buildings with varying slenderness and side ratios under different wind directions. A series of wind tunnel experiments was conducted on six rectangular tall buildings at a scale of 1:300 in a boundary layer wind tunnel. Validated computational fluid dynamics (CFD) models, integrating embedded large eddy simulations for turbulence modelling, were used to generate wind pressure time histories for 15 rectangular tall buildings with varying geometric proportions under seven wind directions (0°, 15°, 30°, 45°, 60°, 75°, and 90°). These CFD data were subsequently used to train a surrogate model based on an attention-based neural network. The neural network surrogate model accurately captured the time-dependent wind pressure distributions over the surfaces of tall buildings with different geometries, achieving an R 2 value of 0.98. It was able to closely replicate and reconstruct complex flow features such as separation, conical vortices, and steep pressure gradients, in good agreement with CFD results. Additionally, statistical measures of pressure coefficients, mean, standard deviation, and extreme values predicted by the neural network showed good agreement with corresponding (unseen) data from wind tunnel experiments. The proposed surrogate model is a promising complementary approach for accurately predicting unsteady wind pressure on tall buildings with varying geometries within a few seconds. • Unsteady wind pressure on tall buildings was predicted using attention-based neural network • Wind tunnel experiments (6 buildings) and numerical modelling (15 buildings) were conducted to generate data • Surrogate model was trained and validated based on numerical modelling data • Neural network accurately reconstructed the time-dependent flow features of the buildings • The neural network predictions are in good agreement with the wind tunnel results