Probabilistic learning approach for the liquid holdup analysis of high-viscosity intermittent flows

Abstract

A Gaussian mixture model (GMM) was implemented to investigate the relationship between the liquid holdup (in various parts of the flow) and the pressure for different experimental realizations of high-viscosity gas–liquid flows. We considered a Newtonian fluid with a constant viscosity of 6 Pa s (600 cP) under a laboratory-controlled temperature. Because the pressure and the holdup do not exhibit a clear-cut relationship in the time domain, a supervised classification algorithm and a “deep” neural network (DNN) were first applied to classify the data points and predict average holdup values. Then, the GMM was applied to determine the holdup in various liquid aggregation structures of the flow as a function of the pressure. The growth rates of the cumulative lengths of the liquid structures (i.e., slug body, mixing front, and liquid film) and the gas bubbles were obtained. The GMM predicted holdup values were in close agreement with the experimental data.