12/05/2024
By Danielle Fretwell

The Francis College of Engineering, Department of Chemical Engineering, invites you to attend a Doctoral Dissertation Proposal defense by Anadi Mondal on: "Predictive Modeling of Entrainment Fraction in Gas-Liquid Annular Flow Using CFD Simulation and Machine Learning"

Candidate Name: Anadi Mondal
Degree: Doctoral
Defense Date: Thursday, December 12, 2024
Time: 11:15 a.m. - 1 p.m.
Location: Southwick Hall, Room 204

Committee:
Advisor: Subash L Sharma, Assistant Professor, Chemical Engineering, UMass Lowell

Committee Members*
Jerome Delhommelle, Associate Professor, Chemistry, UMass Lowell
Dongming Xie, Associate Professor, Chemical Engineering, UMass Lowell

Brief Abstract:
In the Annular gas-liquid two-phase flow regime, determination of the liquid holdup, pressure drop, the onset of dryout, and heat transfer characteristics requires proper knowledge of flow characteristics. A better understanding of entrainment fraction is especially important. However, estimating flow characteristics and entrainment fraction demands large experimental setups associated with high cost and time. To manage this concern, predictive modeling approaches can be utilized to determine flow characteristics and related entrainment fractions accurately.

Firstly, a computational model for the simulation of the annular two-phase flow regime and assessing the various existing models for the entrainment rate has been developed. Computational Fluid Dynamics (CFD) in ANSYS FLUENT has been applied to determine annular flow characteristics including liquid film thickness, film velocity, entrainment rate, deposition rate, and entrainment fraction for various gas-liquid flow conditions in a vertical upward tube. The gas core with droplets was simulated using the Discrete Phase Model (DPM) which is based on the Eulerian-Lagrangian approach. The Eulerian Wall Film (EWF) model was utilized to simulate liquid film on the tube wall. Three different models of Entrainment rate were implemented and assessed through user-defined functions (UDF) in ANSYS. Finally, entrainment for fully developed flow was determined and compared with the experimental data available in the literature. From the simulations, it was obtained that the Bertodano correlation performed best in predicting entrainment fraction and the results were within the ±30 % limit when compared to experimental data.

Besides, a Deep Neural Network (DNN model) has been developed to predict entrainment fraction in annular flow that is applicable for different operating conditions with good accuracy. Usually, empirical correlations are used to predict entrainment fraction. However, these correlations forecast entrainment fractions with large Mean Absolute Percentage Error (MAPE), especially when one correlation is used to predict entrainment fractions for different operating conditions. For instance, existing models considered in this study predict entrainment with MAPE of as much as 120.7%. The neural network is trained and tested for around 1213 entrainment fraction data points extracted from 7 different literature studies. It covers pressure ranges 0.1-20 MPa, tube diameter 5-32 mm with working fluid- air/water, steam/water, and Freon-113. The developed DNN model predicts entrainment fraction with the lowest MAPE compared to the existing models on unseen 155 data from 6 authors that are not used to train or test the model. Furthermore, only 17.61% and 30.81% of data remain outside of ±30% and ±20 limits respectively whereas these values are at least 35.66 % and 52.86% for existing models that are examined for comparison here. After developing an accurate deterministic model, I will focus on uncertainty quantification (UQ) associated with thermal-hydraulics prediction. Uncertainty quantification is an important parameter of the machine learning model which determines the reliability of the developed model. The machine learning model includes mainly two types of uncertainty—epistemic uncertainty and aleatoric uncertainty. I will use a hybrid/ensemble or Bayesian neural network (BNN) for UQ.

All the machine learning models discussed earlier are mainly data-driven. Data-driven modeling doesn’t include physical constraints, equations, or correlations governing output (entrainment fraction or CHF). So, the final part will include the development of a physics-guided machine learning (PGML) model for entrainment fraction by incorporating physical laws. Overall, this study will give a comprehensive understanding of annular flow characteristics in gas-liquid annular flow using computational as well as ML methods.