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

Understanding the liquid-water interaction serves as the basis to credibility of two-phase flow models and
safety of light water reactors. The topic is of researchers’ long interest due to the complexity of underlying
physics. Recently, with growing availability to high performance computing resources, interface tracking
direct numerical simulation becomes an advantage measure to probe the two-phase flow. Resulting
accumulation of high-fidelity numerical data also makes data-driven modeling with machine learning
methods an attractive option to gain insight to the phenomena.

This work presents an interfacial force data-driven modeling framework aims to develop a bubble tracking
direct numerical simulation data-based machine learning drag model for application in Euler-Euler
simulations of bubbly flows. Besides technical demonstration, this work also provides a guidance for DNS
data generation for relevant applications.

The data-driven modeling framework is firstly verified by a benchmark problem, where artificial data is
utilized to make feedforward neural network assimilate drag correlation by Tomiyama et al. (1998). The
obtained model is utilized in a Euler-Euler solver for on-the-fly drag coefficient query. In the test case,
resulting velocity and void fraction distribution by machine learning model is consistent with the reference
model.

Secondly, this work utilized direct numerical simulation bubble tracking data set to form machine learning
drag model for bubbly flow based on Reynolds and Eötvös number. Pseudo-steady state filtering in Frenet
frame is carried out to obtain bubble drag coefficient. The machine learning drag model is examined in a
test case by Wang et al. (1987). Results and suggestions for future works are discussed.