Applied machine learning for adiabatic gas–liquid flow pattern prediction in small diameter circular tubes: Effect of dimensionality reduction

Elham Mollaie, Rasool Mohammadi, Mohammad Ali Akhavan-Behabadi, Behrang Sajadi,
Applied machine learning for adiabatic gas–liquid flow pattern prediction in small diameter circular tubes: Effect of dimensionality reduction,
International Journal of Multiphase Flow,
2025,
105508,
ISSN 0301-9322,
https://doi.org/10.1016/j.ijmultiphaseflow.2025.105508.
(https://www.sciencedirect.com/science/article/pii/S0301932225003830)
Abstract: This study attempts to establish versatile models, based on 30 flow pattern maps available in literature, employing machine learning (ML) methods, within range of database parameters, for adiabatic gas–liquid flow inside small-diameter tubes, from 0.53 to 5.16 mm. Support vector machines (SVM), artificial neural networks (ANN), and histogram-based gradient boosting (HGB) techniques are applied on two separate sets of carefully engineered input features, one with physical dimensional and one with dimensionless parameters, to see if dimensional reduction helps with providing better-performing models. The model training and testing procedure is conducted under a cross-validated study aiming to maximize the performance metric during hyperparameter tuning. The average accuracy of SVM, ANN, and HGB on test sets of data is reported as 0.9284, 0.9240, and 0.9620, respectively based on dimensional features. As for the dimensionless set, in the same order, values of 0.9115, 0.9115, and 0.9583 are obtained, indicating superior performance of HGB, along with acceptable results of ANN and SVM models. ANN models demonstrated faster prediction times than SVM and HGB, which makes ANN models more favorable for high-quantity prediction procedures. HGB models showed more robustness, while the SVM models showed the most prediction uncertainty amongst the models. Also, to visualize the model’s performance, several flow pattern maps are reconstructed with all models. Overall, due to the variety of flow behavior types in the database, employing sets of dimensionless numbers does not secure developing more general models and the performance for different input feature sets is roughly on par.
Keywords: Adiabatic two-phase flow; Flow pattern map; Machine learning; ANN; SVM; HGB; Dimensionless numbers