Application of Supervised Machine Learning Techniques for the Prediction of Pressure Drop and Void Fraction in Gas–Liquid Two-Phase Flow in Horizontal Pipelines

Abstract

Gas–liquid two-phase flow in horizontal pipelines is encountered in a variety of industrial systems, including natural gas gathering networks, chemical reactors, and district heating lines. In such systems the prediction of frictional pressure drop and void fraction is relevant for equipment sizing, control strategies, and safety assessment. Classical approaches are based on mechanistic or empirical correlations that depend on assumptions about flow pattern and interfacial momentum transfer. These approaches can exhibit limited robustness when applied beyond the conditions for which they were developed, especially under wide variations in mass flux, fluid properties, or pipe diameter. Supervised machine learning offers an alternative route that can accommodate nonlinear relations between operating variables and hydraulic responses by learning from combined experimental and numerically generated data. This work examines the application of several supervised learning techniques to the prediction of pressure drop and cross-section-averaged void fraction in horizontal gas–liquid flow. The study considers models ranging from linear regressors to nonlinear kernel methods, ensemble tree algorithms, and feed-forward neural networks, and compares them under a common data-processing workflow. Particular attention is given to the use of non-dimensional inputs, to the incorporation of physically motivated constraints, and to the analysis of model generalization across operating ranges. The results illustrate how data-driven models can complement mechanistic formulations and how error structure, uncertainty representation, and sensitivity analysis can be used to interpret the learned relations without relying on flow-regime classification schemes.