Evaluation of an LES-based Multi-Fidelity framework for wind loading predictions.

Finely resolved Large-Eddy Simulations (LES) can correctly predict wind loading on the surface of high-rise buildings. However, these simulations require substantial computational time. Thus, to allow LES to be routinely used for wind loading predictions, a significant computational speedup is needed.

This work evaluates the use of multi-Fidelity modeling for wind loading predictions to significantly reduce the computational time required by CFD. We aim to combine data at many points from a Low-Fidelity model, namely a coarse LES, with data at a few points from a High-Fidelity model, namely a fine LES, to provide predictions with accuracy close to that of the fine LES at a fraction of the cost. To do so, we build two surrogate models over the discrepancy between the Low- and the High-Fidelity LES, and we use them to correct Low-Fidelity predictions. We also explore how the Low-Fidelity LES resolution affects the predictive ability of the two models by using two different mesh setups for the coarse LES. The results show that both frameworks allow for a reduction in the RMSE of the Low-Fidelity LES predictions with a substantial decrease in the computational cost. Future work will explore a machine learning-based Multi-Fidelity framework to further reduce the predictions RMSE.