A Data-Driven Neural Network Model to Correct Derived Features in a RANS-Based Simulation of the Flow Around a Sharp-Edge Bluff Body

  • Author / Creator
    Shojaee, Seyedamir
  • In this dissertation, a machine-learning method is utilized to enhance the accuracy of wake parameters calculated by Reynolds Averaged Navier Stokes (RANS) k-ω SST model of flow on and around wall-mounted rectangular cylinders. Using high-quality results from Large Eddy Simulation (LES), this correction method at Reynolds number of 2500 saves computational cost and time, while yielding more accurate wake parameters.
    The performance of k-ω SST model to predict the mean wake characteristics and mean global parameters is evaluated in comparison to LES results of the flow around wall-mounted rectangular cylinders. The comparison reveals the inaccuracy of k-ω SST model to predict these characteristics. Different cylinder aspect ratios are considered with depth ratios ranging from 1 to 4 at 0.5 increments. To correct the prediction accuracy of the k-ω SST model, a backpropagation multi-layer perceptron artificial neural network is developed, introduced and tested.
    The qualitative and quantitative comparison of the results of LES and k − ω SST model show that the fundamental properties of flow features over wall-mounted rectangular cylinders, including shear-layer separation, recirculation, and reattachment, are successfully captured by the k−ω SST model. In addition, the k − ω SST model accurately predicts wake topologies, including the tip, base, and horseshoe vortices. However, the global features of the mean flow, mean wake structures, and mean flow features differ significantly from the LES results.
    The neural network was effectively trained using reliable data to improve the accuracy of the mean wake features and global parameters simulated by k − ω SST model. These features and parameters are the reattachment length (XR) behind the cylinder, recirculation length in front of the cylinder, top recirculation length on the top surface of the cylinder, mean drag coefficient, mean lift coefficient, and mean base pressure coefficient. Furthermore, the trained algorithm correctly predicted unobserved data in addition to the given data, which demonstrates the robustness of the developed correction algorithm.

  • Subjects / Keywords
  • Graduation date
    Spring 2023
  • Type of Item
  • Degree
    Master of Science
  • DOI
  • License
    This thesis is made available by the University of Alberta Libraries with permission of the copyright owner solely for non-commercial purposes. This thesis, or any portion thereof, may not otherwise be copied or reproduced without the written consent of the copyright owner, except to the extent permitted by Canadian copyright law.