Abstract Large-eddy simulations (LES) can provide accurate predictions of wind loads on buildings, but their high computational cost, and the need to explore all wind directions with a 10° resolution, limits their use in the design process. Reynolds-averaged Navier–Stokes (RANS) have a low computational cost, but their accuracy can be compromised by the turbulence model and by the model required to retrieve the pressure fluctuations, that ultimately determine the design loads. This study proposes a multi-fidelity machine learning framework that combines computationally efficient RANS, for a large number of wind directions, with more expensive LES, for a small number of wind directions, to provide accurate predictions of the root mean square pressure coefficient at a reasonable computational cost. The training set includes 5 wind directions with a 20° resolution; the test set contains the 5 intermediate wind directions. A bootstrap algorithm, used to generate an ensemble of models, provides confidence intervals that encompass the majority of the LES data for the test directions. These results demonstrate that multi-fidelity machine learning frameworks provide a route to balancing accuracy and computational cost in the prediction of complex turbulent flow quantities.
Paper
Full text
A multi-fidelity machine learning framework to predict wind loads on buildings
Semantic Scholar · Engineering · 2021
Abstract
Abstract Large-eddy simulations (LES) can provide accurate predictions of wind loads on buildings, but their high computational cost, and the need to explore all wind directions with a 10° resolution, limits their use in the design process. Reynolds-averaged Navier–Stokes (RANS) have a low computational cost, but their accuracy can be compromised by the turbulence model and by the model required to retrieve the pressure fluctuations, that ultimately determine the design loads. This study proposes a multi-fidelity machine learning framework that combines computationally efficient RANS, for a large number of wind directions, with more expensive LES, for a small number of wind directions, to provide accurate predictions of the root mean square pressure coefficient at a reasonable computational cost. The training set includes 5 wind directions with a 20° resolution; the test set contains the 5 intermediate wind directions. A bootstrap algorithm, used to generate an ensemble of models, provides confidence intervals that encompass the majority of the LES data for the test directions. These results demonstrate that multi-fidelity machine learning frameworks provide a route to balancing accuracy and computational cost in the prediction of complex turbulent flow quantities.