MACHINE LEARNING-BASED PREDICTION OF INTERFACIAL TENSION IN CO₂-WATER AND CO₂-OIL SYSTEMS

Abstract

Accurate IFT prediction in CO₂–water and CO₂–oil systems play an important role in enhancing the performance of carbon sequestration, enhanced oil recovery (EOR) and subsurface fluid modeling. The accurate experimental procedures require extended periods of time along with restricted system investigation parameters (Zhang et al., 2023). This research utilizes seven machine learning models to predict IFT at different thermodynamic and compositional conditions: Artificial Neural Networks (ANN), Symbolic Regression–Genetic Programming (SR-GP), Decision Trees (DT), Random Forest (RF), Gradient Boosting Regression (GBR), Support Vector Regression (SVR) and XGBoost. The methodology highlighted the development of predictive models that can accommodate the complexity of interfacial phenomena. Carefully curated datasets were used to train and test machine learning algorithms sensitive to the intricate relationship that governs CO₂–water and CO₂–oil systems. The ability of each model to handle non-linear and complex interactions was exhaustively examined to identify the best IFT prediction approach. Model training and testing involved extensive datasets consisting of CO₂–water and CO₂–oil systems defused by various parameters including pressure, temperature, salinity, salt types, oil API gravity and impurities. Standard performance metrics consisting of RMSE, MSE, R², MAPE, AIC and BIC determined the assessment of model accuracy. GBR together with XGBoost and RF proved to be the most accurate ensemble models since they achieved R² (coefficient of determination) values greater than 0.97 while delivering superior predictive capabilities in the evaluation of both systems. The predictive capabilities of ANN models increased remarkably when optimized parameters were applied to the hidden layer structure. The experimental data was verified through KDE plots alongside scatter diagrams which showed excellent correlation between predicted results. The results of sensitivity analyses demonstrated that pressure and temperature highly affect the IFT while salt types and impurities had notable effects. The presented work demonstrates that ML techniques can substitute experimental IFT measurement by providing dependable high-speed scalable options that benefit reservoir modeling and process optimization.