A deep neural network-based multi-fidelity framework for wind loading predictions on low-rise buildings in urban areas.

Recent advances in computational wind engineering show that large eddy simulation (LES) can accurately predict mean, rms, and peak wind loads on buildings. However, LES is still too computationally expensive for widespread commercial use, especially as urban areas expand rapidly. A promising solution is the multi-fidelity (MF) modeling, which reduces computational costs while maintaining high accuracy. This approach combines evaluations of low-fidelity (LF) models at many design points with a limited number of high-fidelity (HF) model evaluations. The goal is to achieve HF-level accuracy at a lower cost by building a surrogate model for the discrepancy between LF and HF models and using this surrogate to correct LF predictions. In this study, we use deep neural networks (DNN) to construct the discrepancy surrogate. Previously, this approach was developed to predict the quantities of interest (QoIs) for high-rise buildings. We extend the model capabilities to low-rise buildings in urban environments. A key enhancement is the integration of a convolutional autoencoder, which extracts features from the urban flowfield. This enables the DNN to effectively capture interactions between wind flow and densely packed buildings, improving the QoI predictions.