Modeling and prediction of slug characteristics utilizing data-driven machine-learning methodology

Slug liquid holdup, translational velocity, and slug length are representative slug characteristics. The precise prediction of these parameters is essential for the accurate calculation of the average liquid holdup and pressure gradient in pipes. Nevertheless, existing correlations have mainly been developed by an empirical approach. Each correlation can only be applied to a restricted fluid property and geometrical condition, preventing people from selecting a suitable one without professional knowledge. In this study, slug characteristics were predicted by several machine learning methodologies with 2590 experimental data to overcome the limitation of selecting the proper model or correlation. The experimental dataset includes -10 degrees <= theta (pipe inclination angle) <= 10 degrees, 0.0258 m <= I.D. (pipe inner diameter) <= 0.0780 m, 0.01 m/s <= vSL (superficial liquid velocity) <= 3.4 m/s, 0.03 m/s <= vSg (superficial gas velocity) <= 20 m/s, 769 kg/m(3) <= rho(L) <= 998 kg/m(3), 1.12 kg/m(3) <= rho(G) <= 7.36 kg/m(3), 0.001 Pa s <= mu(L) <= 4.65 Pa s, and 6 <= Re-M (mixture Reynolds number) <= 863,920. The existing dataset has been preprocessed to pursue better machine learning performance. The random forest results reveal a better training performance and prediction than that of a deep neural network and present a competitive prediction compared with the existing correlations, indicating a great potential of utilizing the data-driven machine learning methodology.