Session: 7.3 - Gas-Liquid flows

Paper Number: 158357

158357 - An Experimental and Data-Based Study of Liquid-Gas Flow in Downward Vertical Tubulars 

Abstract: 

Vertical downward two-phase flow plays a crucial role in various industrial applications, including hydrocarbon injection wells, chemical reactors, and nuclear cooling systems, as such accurate prediction of flow patterns, void fractions, and pressure gradients is essential for optimizing system performance and operational efficiency. This study integrates theoretical and experimental approaches, and predictive capabilities of machine learning models to advance the understanding of vertical downward liquid-gas flow in tubular systems.

An experimental setup was designed comprising a 25-ft vertical flow loop equipped with quick closing valves, differential pressure sensors and high-speed camera. The experiments, conducted with water and air phases, explored a matrix of 15 v_Sg ​ and 5 v_SL​ values. For every test, the flow measurement considers the liquid holdup, pressure gradient, and flow pattern. The results are compared with the predictions from multiple widely used models, such as TUFFP unified, OLGA, and Ansari et al. (1993). In addition, a pretrained regression and classification machine learning models, CatBoost and LGBM respectively, trained using data from 11 published works on vertical downward two-phase flow were validated using experimental data as unseen data, highlighting the potential for machine learning applications. For the pressure gradient prediction, machine learning models were developed using the experimental data.

Experimental observations revealed distinct flow pattern transitions from slug to churn and to annular flows, as v_Sg increased. However, the flow patterns differ from those of upward flow because downward flow is gravity-driven. Void fraction showed a non-linear increase with rising v_Sg​, while pressure gradients decreased, transitioning from gravity-dominated to friction-dominated regions. Higher v_SL values delayed transitions to annular flow, expanding the slug and churn flow regions.

The experimental results were compared with predictions from established physical models (OLGA, Gregory, Ansari, and TUFFP) and those from CatBoost and LGBM. CatBoost exhibited improved accuracy in void fraction prediction, with an average error of 1.28%, outperforming OLGA’s 1.49%. Amongst the physical models, OLGA produces the greatest results, with the TUFFP unified model coming in second. Flow pattern predictions using LightGBM closely matched experimental observations, validating its effectiveness. Finally, gradient boosting demonstrated high accuracy in pressure gradient prediction, achieving an R² of 97%.

This work provides a comprehensive dataset on liquid-gas flow in downward vertical tubulars, a scarce resource in the literature. In addition, it demonstrates that machine learning models provide a robust alternative to traditional mechanistic approaches in predicting complex multiphase flow behaviors. The findings contribute to advancing flow prediction models and offer practical insights for industrial applications requiring accurate modeling of vertical downward two-phase flow systems.

Presenting Author: Oluchi Osuagwu University of Oklahoma

Presenting Author Biography: