|
Abstract
|
One of the important parameters in underground carbon dioxide storage is the CO2-brine interfacial tension (IFT). Although previous studies have estimated this parameter using machine learning methods, many methods have not yet been thoroughly investigated. In this study, the estimation of CO2-brine IFT was addressed using a wide range of decision tree-based regression methods, including regression tree, AdaBoost, XGBoost, Extra Trees, CatBoost, LightGBM, and LSBoost. Initially, outliers were identified and removed using the 3 standard deviation method. Then, the optimal train-to-test data ratio was determined, and based on that, the data were randomly divided into training and testing sets, and the models were developed. According to the results of the present research, the optimized LSBoost method with the Bayesian algorithm, achieving an R2 of 0.9984, exhibited the best performance among the models, while the optimized adaboost method with random search had the weakest performance. Additionally, using kernel density estimation plots, it was determined that, except for adaboost, none of the models showed bias. Further sensitivity analysis using Pearson linear correlation indicated that pressure had the highest impact on CO2-brine IFT values, while temperature had the least impact.
|