11. Data-Driven Machine Learning Applied to Liquid-Liquid Flow Pattern Prediction

The two-phase flow pattern prediction in pipes is a crucial design factor for the energy industry, given its influence on the system’s pressure drop and hold-up. A common approach is to use phenomenological models to predict those parameters as a function of the two-phase flow pattern. As more experimental data became available in the literature, the machine learning methodologies also became an option. In this work, we evaluate the use of data-driven machine learning to predict the liquid-liquid flow pattern transition. The database comprises data from the open literature. Although there is not the same amount of liquid-liquid flow data available, unlike for gas-liquid flow, this study shows that it is possible to predict the liquid-liquid flow patterns using the data-driven approach regardless of the viscosity ratio. Dimensionless parameters derived from the two-phase flow’s governing equations are used to train XGBoost. Four main groups of flow patterns were used in this work: stratified, intermittent, annular, and dispersed. The algorithm’s hyperparameters are tuned using cross-validation and accuracy as the target metric. In addition, a per-flow-pattern-map accuracy is also shown.