Multi-Fidelity Machine Learning Analysis of Wind Patterns Around High-Rise Buildings

Abstract This study presents a multi-fidelity machine learning (ML) approach to predict the fluctuating components of surface pressure coefficients on high-rise buildings across the entire wind rose. Employing an artificial neural network (ANN), the model is trained using a combination of four Reynolds-averaged Navier-Stokes (RANS) simulations at 10°, 20°, 30°, and 40° wind directions, and two large-eddy simulations (LES) at 0° and 45° wind directions. Hyperparameter optimization, model regularization, and feature scaling were conducted to enhance the performance of the trained model and prevent overfitting. Evaluation metrics including root mean square error (RMSE) for the model prediction accuracy study and R2 for the model fit study were utilized; low RMSE and R2 > 0.8 indicated well-trained models. The optimal and regularized model achieved R2 = 0.92 with nine standardized input features. Results demonstrate the effectiveness of the multi-fidelity approach in accurately predicting wind loads across the whole wind rose, offering valuable insights for different city/community planning applications.