Application and assessment of a GPU-based LES method for predicting dynamic wind loads on buildings

Abstract

This study presents the assessment of a fast Large Eddy Simulation method for estimating dynamic wind loads on buildings using a GPU-based CFD software, which produces statistically converged results on a nine-million-cell mesh in approximately 6 hours. The surface pressure distribution of a cuboid building model was validated with experimental data obtained in an atmospheric boundary layer wind tunnel and compared with field measurements. Although due to the applied equidistant Cartesian grid the large gradients near the edges are not fully resolved, good overall agreement was found for the mean and fluctuating pressure distributions (correlation coefficient: 0.90/0.73, FAC2: 0.92/0.98, FB: −/0.06, MG: −/0.95, NMSE: −/0.10, VG: −/1.08). It was shown that the numerical model is able to produce matching turbulent spectra in an intermediate frequency range within the inertial subrange, limited by the domain size and the spatial resolution. Mesh refinement for capturing large gradients as well as for expanding the frequency limits can be achieved by using a GPU with higher VRAM capacity for the simulation. The continuing advancement of the presented model is a promising development for estimating dynamic wind loads on buildings and identifying design problems fast enough for the engineering practice, without high-performance computing.