Development of Machine Learning Framework for Interfacial Force Closures Based on Bubble Tracking Data

This work aims to develop data-driven modeling framework with the aid of machine learning methods and high-fidelity dataset. To gain confidence on the methodology, a bubble drag regression task using artificial dataset is conducted. Result shows FNN’s capability performing non-linear fitting. On the other hand, the sample size test would give sense on model underfitting with same amount of knowledge. Inspired by the previous task, the focus then moved on to utilize DNS bubble tracking dataset for modeling interfacial momentum exchange terms. A novel way to approach interfacial momentum exchange is proposed. Preliminary result reveals the concern of model accuracy on unseen data points. Improvement on model generalization is suggested. Also, further refinement on label formation and data processing should be taken care of. Nonetheless, the potential using high fidelity data and NN to directly model interaction between phases in bubbly flow has been shown.