Methodology for DNS Data-driven Machine Learning Bubble Drag Model and Its Integration to OpenFOAM

This work aims to develop a two-phase DNS data-driven bubble drag model and to implement it into a multiphase flow CFD simulation. To accomplish the goal, a Tensorflow (TF)-OpenFOAM(OF) integration interface has been established. Such an interface is capable of calling and making machine learning model to predict a quantity of interest on the fly. A benchmark case for the bubble drag coefficient is proposed to validate the interface. A Feed forward neural network (FNN) approach was utilized to approximate the drag correlation (Tomiyama et al., 1998) using artificially generated data. Results of the integration showed good consistency in radial void fraction and velocity profiles. As the next step actual DNS bubble tracking datasets are used as a data source (Fang et al., 2017, Cambareri et al., 2019). The data segments where bubble have quasi-stable main-stream velocity were filtered out for drag coefficient calculation. The DNS-informed model predicts bubble drag coefficient by taking bubble Reynolds number (Re) and Eötvös number (Eo) as input to consider the effects from local fluid and bubble shape. The model is applied in a Euler-Euler two-phase flow simulation of a bubbly pipe flow in OF. The required closure terms, except the drag model, utilize the baseline model of Liao et al. (2020) The results of radial void fraction and velocity profiles are discussed and compared to a reference solution with the baseline model.