Integrated Mechanistic Analysis and Machine Learning Prediction of Slug Flow in Oil-Gas-Water Three-Phase Pipelines

Slug flow represents one of the most critical and operationally challenging regimes in oil-gas-water multiphase pipelines. To advance both mechanistic understanding and predictive capability, this study integrates physical analysis with data-driven modeling to elucidate the conditions governing slug formation and to enable its rapid and accurate prediction. A systematic review of existing research is first undertaken to clarify the mechanisms responsible for slug initiation. The influences of gas superficial velocity, liquid velocity, liquid viscosity, liquid surface tension, and the axial component of gravity are examined to characterize their roles in interfacial instability and flow transition. Then, the effects of temperature, total flow rate, water cut, gas-liquid ratio, and pipeline inclination angle are quantitatively assessed, revealing the dominant trends that promote or inhibit slug development. Building on this foundation, a comprehensive three-phase oil-gas-water flow model is constructed. Numerical simulations are performed for 243 operating conditions encompassing a broad range of temperatures, water cuts, gas-liquid ratios, liquid flow rates, and inclination angles. These simulated cases constitute the training dataset for nine machine learning algorithms. To evaluate generalization performance, 108 additional randomly generated operating conditions are predicted, covering temperatures of 80–150°C, water cuts of 40–90%, gas-liquid ratios of 3–30, liquid flow rates of 100–200 t/d, and inclination angles of 5–15. Comparative validation reveals marked differences in predictive accuracy. The BP neural network achieves the highest accuracy, 95%, substantially outperforming XGBoost, 83.3%, Random Forest and Decision Tree, 81.5%, Logistic Regression and Support Vector Machine, 80.6%, K-Nearest Neighbor and Naive Bayes 78.7%, and K-Means, 63%. Overall, the BP neural network demonstrates superior robustness and precision in predicting previously unseen operating conditions, effectively combining the physical consistency of mechanistic modeling with the efficiency and adaptability of machine learning approaches.