Backpropagation Analysis of Swirling Isothermal Dump Combustor Data
Open Access
- Author:
- Werkheiser, Nathan
- Graduate Program:
- Mechanical Engineering
- Degree:
- Master of Science
- Document Type:
- Master Thesis
- Date of Defense:
- March 16, 2022
- Committee Members:
- Rick Ciocci, Professor in Charge/Director of Graduate Studies
Brian Allen Maicke, Thesis Advisor/Co-Advisor
Anilchandra Attaluri, Committee Member
Seth Wolpert, Committee Member - Keywords:
- Neural Networks
Backpropagation
fluid flow
CFD
Swirling Flow
Activation Functions - Abstract:
- Swirling flow is a difficult topic to try to simulate accurately. Many turbulence models like k-ε or k-ω run into multiple problems and inaccuracies when trying to simulate a turbulent flow that also has swirl. Although Computational Fluid Dynamics (CFD) is currently a field dominated by numerical analysis, advancements in machine learning show promise as a tool to develop full scale methodologies to solve complex or ineffective simulations, or to better initialize a simulation with artificial intelligence models. This paper focuses on analyzing experimental data that was gathered from an Isothermal Dump Combustor. A swirler is used to induce a swirl to the flow, and the resulting velocities are measured using a Laser Doppler Vibrometer (LDV). The LDV then gathered data at multiple points along the length of the combustion chamber, analyzing across the entire positive radial coordinates. The data used to train the neural network were the axial, radial, and tangential velocities, Reynold’s shear stresses, and the turbulent kinetic energy. The neural network utilized a feed forward network, only utilizing backpropagation. This model explored multiple activation functions to explore the accuracy, training cycles, and gradients of the data in comparison with the experimental results. After training each of these models, the data was plotted to compare to the experimental data. The four activation functions that were explored were tanh, RELU, SELU, and swish Overall, the study showed promising results for neural networks in analyzing swirling flow. The tanh activation function had the best RMS error and showed consistent contours. Other saturated models, like the swish and sigmoid also had much smoother gradients and transitions on the contour plots. Non-saturated models, like Rectified Linear Unit (RELU) and Scaled Exponential Linear Unit (SELU), also showed reasonable accuracy but exhibited jagged contour plots.