Session: 10-09-01: Multiphase Flows and Applications

Paper Number: 111411

111411 - A Data-Driven Approach for Predicting the Onset of Entrainment in Two-Fluid Stratified Systems During Selective Withdrawal Process Using Machine Learning Techniques 

Selective withdrawal is a desired phenomenon in transferring oil from large caverns in the US Strategic Petroleum Reserve (SPR). The entrainment of oil during withdrawal poses a risk of contaminating the environment. Other applications of selective withdrawal include the coating of microparticles and the fabrication of thin glass fibers. This process is also important for local drug delivery and is used in many medical settings. Physics-based mechanistic models are often used to predict the behavior of complex phenomena, such as the onset of entrainment during selective withdrawal processes. But these models have significant uncertainties due to the experimental conditions, variability of dimensions, and flow conditions. The predictive relationships found in the literature are limited to specific flow regimes. In contrast, data-driven machine learning models can bypass these aforementioned issues and provide a quick and efficient generalized solution to these complex problems. 

This study explores the effectiveness of machine learning methodologies, specifically Gaussian Process Regression (GPR) and Neural Networks (NN), in predicting the onset of entrainment during selective withdrawal in two-fluid stratified systems. Gaussian processes are non-parametric models that perform well in highly non-linear data and with small data sets with high noise. GPR takes a Bayesian approach, which means it seeks to model the full distribution of possible probability functions that could explain the data. This allows the model to capture uncertainty in small datasets and make more robust predictions. On the other hand, Neural networks are considered universal function approximators because they can approximate any continuous function to arbitrary accuracy, given enough nodes and layers in the network. This is because multiple layers allow the model to learn hierarchical representations of the input data, which can capture increasingly complex and abstract features.

In this study, the Machine Learning (ML) models were trained using experimental data obtained from various sources in the literature. The collected data included those from Lubin et al. and Sabbir et al., which were from the low Ohnesorge number (Oh) regime where surface tension and inertia effects were dominant over viscous effects, and from Cohen et al. and Eggers et al., which were from the high Oh regime where viscous effects were dominant over surface tension and inertia effects. To ensure the generalizability of the ML models, appropriate non-dimensional parameters such as Froude number, Weber number, Capillary number, Reynolds number, and Bond number were identified and used as model inputs. These non-dimensional flow parameters inherently provided information regarding the flow regime based on the appropriate length scales. To further enhance the efficiency and precision of the prediction of the onset of entrainment, dimensionality reduction techniques, such as Genetic Algorithm (GA), were applied to circumvent the problem of multicollinearity between the non-dimensional features and to identify the most significant ones. Finally, the predicted results from the ML models were compared with the predictions of the physics-based mechanistic models at various flow regimes, and a comparative performance analysis was conducted.

The preliminary findings of this study indicate that both the Gaussian Process Regression (GPR) and Neural Network (NN) models exhibit excellent predictive capabilities.  This is evidenced by the observation of lower values of the Root Mean Squared Error (RMSE) and Mean Absolute Percentage Error (MAPE) with GPR showing slightly better performance than NN models. Moreover, both models surpass the predictive capability of all physics-based models across various flow regimes. However, predictions made by the models will hold only if the underlying physical regime remains unchanged. Any variation in the physical behavior of the flow regime may lead to further deviation in the results.

Machine learning models can unify various physics-based models based on available training data at different flow regimes. Overall, our study highlights the potential of machine learning techniques in solving complex problems and offers a novel approach to predicting the onset of entrainment in two-fluid stratified systems.

Presenting Author: Sabbir Hassan Texas Tech University

Presenting Author Biography: Sabbir Hassan is a Ph.D. candidate and Graduate Part-time Instructor in the Department of Mechanical Engineering at Texas Tech University. Prior to that, he earned his Bachelor of Science (B.Sc.) from Bangladesh University of Engineering and Technology in Dhaka, Bangladesh. His current research focuses on developing a predictive model for fluid entrainment during selective withdrawal, utilizing experimental data to observe the entrainment of fluids and extrapolating these findings to different flow regimes. His expertise in fluid dynamics has been recognized recently as a publication titled “Predicting transition from selective withdrawal to entrainment in two-fluid stratified systems” in the prestigious journal Physical Review E. In addition, he is interested in the application of machine-learning techniques to physics-based fluid phenomena. Through his work, he aims to enhance the extrapolation capabilities of these models, enabling a deeper understanding of complex fluid systems.

Authors:

Sabbir Hassan Texas Tech University
Darryl James Texas Tech University