Machine learning based models for pressure drop estimation of two-phase adiabatic air-water flow in micro-finned tubes: Determination of the most promising dimensionless feature set

The present study is focused on determining the most promising set of dimensionless features and the optimal machine learning algorithm that can be employed for data-driven frictional pressure drop estimation of water (single-phase) and air-water mixture (two-phase) flow in micro-finned horizontal tubes. Accordingly, an experimental activity is first conducted, in which the frictional pressure drop of both water and air-water flows, at various flow conditions, is measured. Next, machine learning based pipelines are developed, in which dimensionless parameters are provided as features and the friction factor (for the single-phase case) and the two-phase flow multipliers (for the two-phase case) are considered as the targets. Next, the feature selection procedure is performed, in which the most promising set of features, while employing a benchmark algorithm, is determined. An algorithm optimization procedure is then performed in order to choose the most suitable algorithm (and the corresponding tuning parameters) that lead to the highest possible accuracy. Moreover, the state-of-the-art physical models are implemented and the corresponding accuracy, while being applied to the experimentally obtained dataset, is determined. It is demonstrated that only 5 dimensionless features are selected (among 23 provided features) in the obtained pipeline developed for the estimation of the two-phase gas multiplier (in the extraction procedure of which, the single-phase friction factors are determined only using the Reynolds number and two geometrical parameters). Therefore, the latter procedures notably reduce the complexity of the model, while providing a higher accuracy (MARD of 6.72% and 7.05% on the training and test sets respectively) compared to the one achieved using the most promising available physical model (MARD of 15.21%). Finally, through implementing the forward feature combination strategy on the optimal pipeline, the contribution of each feature to the achieved accuracy is shown and the trade-off between the model's complexity (number of features) and the obtained accuracy is presented. Thus, the latter step provides the possibility of utilizing an even inferior number of features, while achieving an acceptable accuracy. Moreover, since these pipelines will be made publicly accessible, the implemented models also offer a higher reproducibility and ease of use.