Using an Artificial Neural Network to Predict Drag, Lift, and Torque Coefficients of an Axisymmetric Ellipsoid in a Viscous Fluid




Hinojosa, Daniel Andrew

Journal Title

Journal ISSN

Volume Title



Computational Fluid Dynamics (CFD) simulations on spheres have been the dominant method of late for solving gas-solid force interactions. However, the spherical based solutions contain inherent errors when simulating systems with non-spherical particles. Improvements in artificial neural networks (ANN) combined with advancements in computer processing speeds and memory capabilities make machine learning an arguably better choice for analyzing interactions in fluidized-particulate systems. A Back Propagation Neural Network (BPNN) is evaluated to predict drag, lift, and torque coefficients of non-spherical particles in gas-solid two-phase flow. This effort aims to reduce time, financial, and computational burdens. The predictions are compared to calculated results based on a finite difference method in-house code (FDM ihc). A hyper-parameter tuning code and k-fold validation are conducted to help design the model architecture. The features fed to the three-layer framework include the incident angle (uD835uDF03), aspect ratio (uD835uDEFD), and Reynolds number (uD835uDC45uD835uDC52). The percent error of the normalized drag coefficient and the relative error of the lift and torque coefficients is < 5% for more than 99% of the test set. The maximum mean square error (MSE) is on the order of 10-4 and the maximum mean absolute error (MAE) is on the order of 10-3. The combination of low relative errors and high-performance metrics in this thesis supports the new wave application of artificial intelligence to replace lower-level methods to study interactions in fluid-solid systems with less time, reduced monetary commitment, and lower computational requirements.Additionally, this paper develops and validates a set of equations to calculate the drag coefficient for an ellipsoid in a viscous Newtonian fluid. The proposed equations whose functional form are polynomial setups depend on the uD835uDC45uD835uDC52, aspect ratio, and the incident angle between the particle and the incoming fluid flow. The relationships between these features to obtain the calculations are mimicked after those by published literature. Curve fitting methods are employed to determine the coefficients in the correlations. The results show very good agreement with the FDM ihc data and the ANN predictions in this paper and previous works.



Computational Fluid Dynamics, Computer processing speeds, Predictions, Artificial neural networks, Monetary commitment



Mechanical Engineering