Simplifying assumptions and empirical closure relations are often required in existing two-phase flow modeling based on first-principle equations, hence limiting its prediction accuracy and in some instances compromising safety and productivity. State-of-the-art models used in the industry still include correlations that were developed in the sixties, whose prediction performances are at best acceptable. To better improve the prediction accuracy and encompass all pipe inclinations and flow patterns, we propose in this paper an artificial neural network (ANN)-based model for steady-state two-phase flow liquid holdup estimation in pipes. Deriving the best input combination among a large reservoir of dimensionless Π groups with various fluid properties, pipe characteristics, and operating conditions is a laborious trial-and-error procedure. Thus, a self-adaptive genetic algorithm (GA) is proposed in this work to both ease the computational complexity associated with finding the elite ANN model and lead to the best prediction accuracy of the liquid holdup. The proposed approach was implemented using the Stanford multiphase flow database (SMFD), chosen for being among the largest and most complete databases in the literature. The performance of the proposed approach was further compared to that of two prominent models, namely a standard empirical correlation-based model and a mechanistic model. The obtained results along with the comparison analysis confirmed the enhanced accuracy of the proposed approach in predicting liquid holdup for all pipe inclinations and fluid flow patterns.

Real-time monitoring of pressure and flow in multiphase flow applications is a critical problem given its economic and safety impacts. Using physics-based models has long been computationally expensive due to the spatial–temporal dependency of the variables and the nonlinear nature of the governing equations. This paper proposes a new reduced-order modeling approach for transient gas–liquid flow in pipes. In the proposed approach, artificial neural networks (ANNs) are considered to predict holdup and pressure drop at steady-state from which properties of the two-phase mixture are derived. The dynamic response of the mixture is then estimated using a dissipative distributed-parameter model. The proposed approach encompasses all pipe inclination angles and flow conditions, does not require a spatial discretization of the pipe, and is numerically stable. To validate our model, we compared its dynamic response to that of OLGA©, the leading multiphase flow dynamic simulator. The obtained results showed a good agreement between both models under different pipe inclinations and various levels of gas volume fractions (GVF). In addition, the proposed model reduced the computational time by four- to sixfolds compared to OLGA©. The above attribute makes it ideal for real-time monitoring and fluid flow control applications.