An Improved Model for Gas-Liquid Flow Pattern Prediction Based on Machine Learning

Mask, Gene (University of Oklahoma) | Wu, Xingru (University of Oklahoma) | Ling, Kegang (University of North Dakota)

OnePetro 

Abstract

Flow pattern of a multi-phase flow refers to the spatial distribution of the phase along transport conduit when liquid and gas flow simultaneously. The determination of flow patterns is a fundamental problem in two-phase flow analysis, and an accurate model for gas-liquid flow pattern prediction is critical for any multiphase flow characterization as the model is used in many applications in petroleum engineering. The objective of this study is to present a new model based on machine learning techniques and more than 8000 laboratory multi-phase flow tests.

The flow pattern is affected by fluid properties, in-situ flow rates of liquid and gas, and flow conduit geometry and mechanical properties. Laboratory data since 1950s have been collected and more than 8000 data points had been obtained. However, the actual flow conditions are significantly different with any laboratory settings. Therefore, several dimensionless variables are derived to characterize these data points first. Then machine learning techniques were applied on these dimensionless variables to develop the flow pattern prediction models. Applying hydraulic fundamentals and dimensional analysis, we developed dimensionless numbers to reduce number of freedom dimensions. These dimensionless variables are easy to use for upscaling and have physical meanings. We converted the collected data from actual laboratory measurement to the variations of these dimensionless variables. Machine learning techniques on the dimensionless variable significantly improved their predictive accuracy. Currently the best matching on these laboratory data was about 80% using the most recently developed semi-analytical models. Using machine learning techniques, we improved the matching quality to more than 90% on the experimental data.

This paper applies machine learning techniques on flow pattern prediction, which has tremendous practical usages and scientific merits. The developed model is better than current existing semi-analytical or classical correlations in matching the laboratory database.