Smart Algorithms for Determination of Interfacial Tension (IFT) between Injected Gas and Crude Oil-Applicable to EOR Projects

Rampant move towards the energy transition landscape has propelled the need to reduce and manage greenhouse gas (GHG) emissions. To this end, it is pivotal to explore cost-effective avenues to utilize GHG and flue gases that can lead to achieving a reduction in emissions target and provide EOR opportunities. Accordingly, the injection of these gases in oil reservoirs allows for hitting two targets with one arrow. Nevertheless, the success of these projects is dependent on accurate IFT estimates between injected gas and resident hydrocarbons. Therefore, this paper evaluates the utility of machine learning algorithms for the estimation of IFT between a mixture of CO₂:N₂ and straight-chain alkanes for variable temperature/pressure conditions. A systematic approach is adopted to implement predictive models for IFT prediction by utilizing over 250 experimental data points from the available literature. A comprehensive statistical analysis is performed to achieve model generalization capabilities and improve control over the most relevant input parameters. Consequently, IFT is demarcated as a function of four readily available inputs with pressure and temperature being the most obvious, while carbon number and flue gas mole fraction is also incorporated as critical parameters. Various smart approaches in this work are proposed through the development of an IFT predictor using Artificial Neural Network (ANN), Support Vector Regression (SVR), and Extreme Gradient Boosting (XGBoost) algorithms. Machine learning model training is enhanced using a K-fold cross-validation technique combined with the exhaustive grid search algorithm. Post-training, the developed models are tested for reliability using blind datasets reserved for this purpose. A fair comparison between model efficiency is ensured by using an in-depth error analysis schema that includes various metrics like the correlation of determination, average error analysis, graphical error analysis, and scatter plots. This allows a relative ranking system to be generated that weighs various factors to classify one model as the most efficient. For the IFT prediction problem, it was found that the neural network was aptly able to yield high efficiency and low errors. This stems from the way ANNs function to map the non-linearity relationship between carbon number and the IFT. It was also observed that enhancement of the intelligent model training through multiple techniques resulted in optimized hyperparameters/parameters.