THEORETICAL AND EXPERIMENTAL ANALYSIS OF DOWNWARD LIQUID-GAS FLOW IN VERTICAL TUBULARS

Abstract

Vertical downward two-phase flow is a critical phenomenon in various industrial applications, hydrocarbon injection wells, chemical reactors, and nuclear cooling systems. The accurate prediction of flow patterns, void fractions, and pressure gradients is vital for optimizing operational efficiency in such flows. This study focuses on the theoretical and experimental analysis of vertical downward liquid-gas flow in tubular systems.Firstly, a machine learning model was developed to predict flow pattern and void fraction using data from 11 published works on vertical downward two-phase flow. Data points, including superficial gas and liquid velocities (vSg and vSL), void fractions and flow pattern were gathered from various studies to build a robust training set. Dimensionless features like Reynolds and Weber numbers of both liquid and gas phases were utilized, along with several machine learning models. CatBoost and LightGBM emerged as the best-performing algorithms for void fraction and flow pattern prediction, respectively. This model served as a foundation for later comparisons with experimental observations, highlighting the potential for machine learning applications. An experimental setup was designed, consisting of a 25-ft vertical flow loop equipped with differential pressure sensors and a camera to capture flow behavior with a test matrix of 15 vSg and 5 vSL values. The experiments were conducted using water and air as the liquid and gas phases to investigate the downward flow dynamics. Images were analyzed to observe flow patterns, along with void fractions and pressure gradients at different flow conditions. The experimental observations revealed distinct flow pattern transitions, including slug, churn, and annular flow, as gas velocity increased. Void fraction was found to increase non-uniformly with rising vSg, while pressure gradients generally decreased with increasing vSg, moving from gravity dominated to friction dominated regions. Higher vSL values influenced the flow patterns by expanding the slug and churn flow regions and delaying transitions to annular flow. The performances of established physical models (OLGA, Gregory, Ansari, and TUFFP) were compared with experimental data. The machine learning models developed, LightGBM and CatBoost, showed improved accuracy in predicting flow patterns and void fractions over traditional models. The results showed that the CatBoost model slightly outperformed traditional models in predicting void fraction, with an average error of 1.28% compared to OLGA’s 1.49%. Pressure gradient predictions were more challenging, with OLGA performing best at 82.49% relative error. Flow pattern observations aided in validating machine learning predictions. LGBM performed well in predicting the flow patterns of experimental data. This study concludes that machine learning models provide a robust alternative to traditional mechanistic models in predicting complex multiphase flow behaviors. The work contributes to the broader understanding of vertical downward multiphase flow and offers practical recommendations for improving flow prediction models in industrial applications.