11/07/2025
By Danielle Fretwell

The Francis College of Engineering, Department of Energy Engineering - Nuclear, invites you to attend a Doctoral Dissertation 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: Tuesday, November 18, 2025
Time: 2 - 3:30 p.m.
Location: Perry 115

Committee:

  • Advisor: Stephen T. Lam, Assistant Professor, Chemical Engineering, UMass Lowell
  • Co-Advisor: Subash L Sharma, Associate Professor, Nuclear Engineering and Radiation Science, MST
  •  Dongming Xie, Associate Professor, Chemical Engineering, UMass Lowell
  • Jerome Delhommelle, Professor, Chemistry, UMass Lowell

Abstract:
Two-phase flow and boiling heat transfer are central to the design, efficiency, and safety of thermal-hydraulic systems such as oil and gas transport, boilers, steam generators, heat exchangers, chemical separation processes, and nuclear reactors. Among these, gas–liquid annular flow is characterized by a liquid film along the wall and a gas core consisting of continuous gas and liquid droplets. Accurate prediction of entrainment fraction (fraction of liquid flows as droplets in gas core) in annular flow is essential for determining liquid holdup, pressure drop, flow transitions, and dryout, all of which govern system performance and safety. This thesis develops Computational Fluid Dynamics (CFD), machine learning (ML), physics-informed ML (PIMLAF), and Bayesian models to improve the prediction of entrainment fraction and critical heat flux (CHF) in annular flow.

Initially, CFD simulations were performed in ANSYS FLUENT using the Eulerian Wall Film (EWF) and Discrete Phase Model (DPM) to simulate annular flow characteristics. Three entrainment models were implemented through user-defined functions, and the Bertodano correlation showed the best agreement with experimental entrainment fraction within ±30%. Next, a Deep Neural Network (DNN) was developed to predict entrainment fraction using 1,213 data points across wide operating conditions and fluid types. The DNN significantly reduced the mean absolute percentage error (MAPE) compared to empirical correlations. Further, to enhance model interpretability and robustness, ensemble ML models—Random Forest (RF) and Gradient Boosting Regression (GBR)—were trained using five non-dimensional input parameters and 1,628 data points. These models achieved the lowest RMSE (root mean square error), MAE (mean absolute error), and MAPE with only 15–20% of unseen data points lying outside the ±30% error band, validating their predictive reliability with an uncertainty quantification.

To incorporate physical knowledge, PIMLAF was proposed by coupling empirical correlations with ML residual learning. The hybrid RF + Cioncolini &Thome model showed superior performance on the extrapolated dataset, with a good prediction accuracy on the test and unseen dataset, demonstrating enhanced generalization and physical consistency.

Finally, a Bayesian Neural Network (BNN) was developed using the NRC (Nuclear Regulatory Commission) CHF database (24,579 data points) to predict CHF and quantify uncertainty. The BNN achieved an MAPE of 7.6% and an R² of 0.986, capturing 94.8% of test data within the 95% confidence interval—outperforming the traditional look-up table. Overall, this research establishes a comprehensive CFD, data-driven, and physics-informed modeling framework for accurate and uncertainty-aware prediction in two-phase thermal-hydraulic systems.