2022 journal article

Development of machine learning framework for interface force closures based on bubble tracking data

NUCLEAR ENGINEERING AND DESIGN, 399.

By: C. Tai, I. Evdokimov, F. Schlegel, I. Bolotnov & D. Lucas

author keywords: Direct numerical simulation; Data-driven modeling; Interfacial force modeling; Machine learning; Eulerian-Eulerian approach
Source: Web Of Science
Added: November 28, 2022

• A data-driven modeling (DDM) framework is established for Eulerian-Eulerian approach. • The implementation of the DDM framework is verified for an artificial problem. • Bubble tracking data set is filtered in Frenet frame to obtain bubble drag coefficients. • The performance of the data-driven drag model is discussed using a pipe flow case study. Interfacial force closures in the two-fluid model play a critical role for the predictive capabilities of void fraction distribution. However, the practices of interfacial force modeling have long been challenged by the inherent physical complexity of the two-phase flows. The rapidly expanding computational capabilities in the recent years have made high-fidelity data from the interface-captured direct numerical simulation become more available, and hence potential for data-driven interfacial force modeling has prevailed. In this work, we established a data-driven modeling framework integrated to the HZDR multiphase Eulerian-Eulerian framework for computational fluid dynamics simulations. The data-driven framework is verified in a benchmark problem, where a feedforward neural network managed to capture the non-linear mapping between bubble Reynolds number and drag coefficient and reproduce the void distribution resulting from the baseline model in the test case. The second focus is on utilizing the bubble tracking data set to form a closure for the bubble drag in the turbulent bubbly flow, in which the drag coefficient is set to be correlated with the bubble Reynolds number and the Eötvös number. Pseudo-steady state filtering in the Frenet Frame was carried out to obtain the drag coefficient from the turbulent bubbly flow data. The performance of the data-driven drag model is also examined through a case study, where improvement of model’s prediction near-wall is regarded necessary. Discussion and further plans of investigation are provided.