Deep Learning-Based Eddy Viscosity Modeling for Improved RANS Simulations of Wind Pressures on Bluff Bodies

Abstract

Accurate prediction of wind pressures on buildings is crucial for designing safe and efficient structures. Existing computational methods, like Reynolds-averaged Navier-Stokes (RANS) simulations, often fail to predict pressures accurately in separation zones. This study proposes a novel deep-learning methodology to enhance the accuracy and performance of eddy viscosity modeling within RANS turbulence closures, particularly improving predictions for bluff body aerodynamics. A deep learning model, trained on large eddy simulation (LES) data for various bluff body geometries, including a flat-roof building and forward/backward facing steps, was used to adjust eddy viscosity in RANS equations. The results show that incorporating the machine learning-predicted eddy viscosity significantly improves agreement with LES results and experimental data, particularly in the separation bubble and shear layer. The deep learning model employed a neural network architecture with four hidden layers, 32 neurons, and tanh activation functions, trained using the Adam optimizer with a learning rate of 0.001. The training data consisted of LES simulations for forward/backward facing steps with width-to-height ratios ranging from 0.2 to 6. The study reveals that the machine learning model achieves a balance in eddy viscosity that delays flow reattachment, leading to more accurate pressure and velocity predictions than traditional turbulence closures like k-ω SST and k-ε. A sensitivity analysis demonstrated the pivotal role of eddy viscosity in governing flow separation, reattachment, and pressure distributions. Additionally, the investigation underscores the disparity in eddy viscosity values between RANS and LES models, highlighting the need for enhanced turbulence modeling. The findings presented in this paper offer substantive insights that can inform the advancement of more dependable computational methodologies tailored for engineering applications, encompassing wind load considerations for structural design and the intricate dynamics of unsteady aerodynamic phenomena.

Keywords

• wind pressures
• deep learning
• eddy viscosity modeling
• rans simulations
• separation bubble
• shear layer
• bluff body aerodynamics