A neural network-based multi-fidelity modeling approach for large-eddy simulations with application to wind loading predictions

Recent advancements in computational wind engineering have demonstrated the potential of Large-Eddy Simulations (LESs) as a design tool for wind loading predictions. However, the computational cost associated with finely resolved LESs capable of capturing peak pressures can become a limiting factor when the analysis requires evaluating the wind loads for all wind directions. The same limitation arises in other LES applications that require predictions across a parameter space. This study proposes a Multi-Fidelity (MF) framework for LES that leverages Neural Networks (NN). A MF data set selection strategy and a MF loss function are proposed to obtain NNs that provide optimal MF model performance. The framework is applied to predict wind loading on a high-rise building across the entire wind rose. Data from 10 low-resolution LESs is combined with data from 5 high-resolution LESs, to provide MF predictions for the point-wise mean, rms, and peak pressure coefficients across the entire wind rose. The MF framework consistently reduces the error associated with the low-resolution LESs for each wind direction, achieving accuracy close to the high-resolution LESs at half of the computational cost. The MF model error averaged across all building facades is consistently lower than the error of the low-resolution model for every wind direction and Quantity of Interest (QoI). The framework can be readily applied to other LES applications.