The use of machine learning methods for fast estimation of CO2-brine interfacial tension: A comparative study

The CO2-brine interfacial tension (IFT) is key to designing the CO2 injection into underground saline aquifers in order to reduce CO2 and slow global temperature increase. Laboratory measurement of CO2-brine IFT is usually time-consuming and requires an expensive experimental apparatus and sophisticated interpretation procedures. This paper presents comprehensive comparisons of nine state-of-the-art machine learning methods for fast estimation of CO2-brine IFT. Results show that: i) the extreme gradient boosting (XGBoost) and gradient boosting decision tree (GBDT) exhibit the strongest robustness and generalization capability, which can be applied for accurate and fast estimation of CO2-brine IFT in practice; ii) the Gaussian process regression (GPR) is associated with the issue of overfitting and therefore may not be reliable for further application; and iii) the other six methods (support vector machine, multi-player perception, kernel ridge regression, classifier and regression tree, random forest, Adaboost) have comparable accuracies in predicting the "unseen" data although noticeable variations in the correlation accuracies are observed for different methods in the training stage.