Session: 10-10-03: Industrial Flows - III

Paper Number: 69010

Start Time: Tuesday, 11:05 AM

69010 - Deep Learning for Drag Coefficient Predictions of Spherical and Non-Spherical Particles 

The tracking of a particle or determining the distribution of multiple particles is challenging for numerical studies about the particle-laden flows, like sand ingestion inside gas turbine engines, fluidized bed for combustion and chemical processing, and range prediction of volcano debris after the eruption. Simulation of such problems using Eulerian Multiphase Flow Model, DPM (Discrete Phase Model) or DEM (Discrete Element Model) where surface forces on particles due to fluid are estimated using drag modeling using the local flow variables. The drag force( is mainly governed by the drag coefficient(C_D) which largely depends on the Reynolds number and shape factors of the particle. Although drag coefficients for spheric particles have been well developed, there is a lack of credible the drag coefficient model for non-spheric particles. As the drag model for non-spherical particles required in a particle-laden flow is not fully established, flow simulation could cover a wide range of sphericities of the particles become very challenging.

Traditional approaches in determining the drag coefficients include experimental measurement and numerical simulations, including Direct Numerical Simulation, which are expensive and time consuming. In this paper, a data-driven model was adopted to develop a relatively more general model for the prediction of drag coefficients. Most of the existing studies have considered sphericities and Reynolds numbers within limited parameter ranges. This study focuses on developing an artificial neural network model by using a large number of available experimental data for a wide range of sphericities (0.034-1), density ratios (0.0005-0.491), and Reynold numbers (0.002-79432).  Available experimental and DNS data for particles of various sizes and materials tested against liquid and gas are identified to correlate the drag coefficient. Three different machine learning algorithms, Random Forest, Gradient Boosting, and a Deep Neural Network (DNN), are trained and evaluated. The neural network results were compared to the experimental results and to select numerical correlations.

Using the available data sets examined in this study, it was found that the simple models, Random forest and XGBoost, can predict more accurate drag coefficients compared to numerical correlations without changes in the default parameters. On the other hand, the DNN model can accurately predict  in a wide-open range of sphericities and Re numbers. The DNN outperforms numerical correlations proposed in previous studies for most of the sphericities, showing the effectiveness of DL models for the prediction of . The DNN model particularly performs better than all the other methods and algorithms for most of the studied sphericities (0.36-1). The DNN predictions were also compared with the numeric correlation used in the MFIX software, and the predictions were more accurate for all the studied sphericities.

Presenting Author: Pratik Mahyawansi Florida International University

Authors:

Pratik Mahyawansi Florida International University
Cheng-Xian Lin Florida International University
Shu-Ching Chen Florida International University