Modeling the Interfacial Dynamics of Carbonated Water-Crude Oil Systems Using Intelligent Boosting Algorithms

Summary Managing carbon dioxide (CO2) emissions in enhanced oil recovery (EOR) remains a crucial global issue. Solutions to store CO2 and boost fossil fuel production are in demand. Dissolving CO2 in water improves the sweep efficiency of EOR and reduces atmospheric CO2. However, experimentally studying these processes is challenging, highlighting the need for effective modeling. In this study, we explore the crude oil-carbonated water interfacial tension (IFT) (one of the essential factors for EOR effectiveness) using four machine learning techniques—light-gradient boosting machine (LightGBM), extreme gradient boosting (XGBoost), adaptive boosting decision tree (AdaBoost-DT), and gradient boosting with categorical features support (CatBoost). We used a data set of 5,279 measurements to develop comprehensive predictive models, accounting for variables such as pressure, temperature, time, salinity, salt type, and oil composition. We tested two models with different input parameters (Model 1: 7 inputs; Model 2: 10 inputs). Both statistical and graphical evaluations demonstrated that all algorithms provided accurate predictions, with CatBoost for Model 2 achieving the best performance, with an average absolute percent relative error (AAPRE) of 1.10%, root mean square error (RMSE) of 0.303, and correlation coefficient (R²) of 0.9953. CatBoost in Model 1 also showed the highest accuracy with AAPRE, RMSE, and R2 values of 1.25%, 0.330, and 0.9944, respectively. Cumulative frequency analysis showed that CatBoost predicted nearly 91% of the data with less than 3% error. Moreover, trend analysis revealed that IFT reduction over time is influenced by naturally occurring surfactants within the crude oil. Furthermore, the optimized CatBoost model performed with extremely high accuracy in aligning with this variation trend. Outlier detection has proved that most data points fall within statistically acceptable parameters, and the databank we used is statistically reliable. The results of this study are widely applicable to EOR procedures, leading to improved oil extraction efficiency and reduced operational costs, thus making the process more reliable and effective.