On March 23, 2018, Yang Liu successfully defended his PhD dissertation, Development of a Data-Driven Analysis Framework for Boiling Problems with Multiphase-CFD Solver. Yang’s committee consisted of his advisor, Nam Dinh, and members, Igor Bolotnov, J. Michael Doster, and Ralph Smith.

Abstract

LIU, YANG. Development of a Data-Driven Analysis Framework for Boiling Problems with Multiphase-CFD Solver. (Under the direction of Prof. Nam Dinh.)

Flow boiling is a highly efficient heat transfer regime, which is used for thermal management in various engineered systems. Among the modeling tools for boiling, the Multiphase Computational Fluid Dynamics (MCFD) solver based on Eulerian-Eulerian two-fluid model has demonstrated its potential in solving boiling problems for industrial applications. On the other hand, in two-fluid model, closure relations are needed to make the two-fluid conservation equations solvable. Such relations, usually empirical or semi-empirical correlations, bring model form uncertainty and model parameter uncertainty to the MCFD solver. A still open issue for MCFD is that such uncertainties can be significant and are still not well quantified, thus undermining the predictive capability of the solver.

This dissertation presents a data-driven analysis framework to address this open issue. The framework aims to leverage state of the art statistical methods and the increasingly affluent boiling data, from both high-resolution experimental measurements and high-fidelity simulations, to 1). perform validation and uncertainty quantification (VUQ) for the MCFD solver based on all available datasets; 2). develop data-driven closure relations based on deep neural networks for the MCFD solver that has the better predictive capability. Three major products are developed within the framework.

First, a boiling data processing and storage procedure is developed for high-resolution experiments and high-fidelity simulations. The extracted data are stored in a structured manner to ensure the flexibility for multipurpose usage. Second, a comprehensive validation and uncertainty quantification (VUQ) procedure is developed for the MCFD solver. The procedure quantifies the uncertainties of MCFD solver predictions using Bayesian inference; then calculates validation metrics that quantitatively measuring the agreement between experimental measurement and obtained prediction uncertainties. Last, a study of new boiling closure relation development based on deep learning is performed. The deep feedforward networks trained by high fidelity boiling simulation data are found to be capable of predicting wall boiling heat transfer behavior with good accuracy.