Cloud point pressure estimation of gas-soluble chemicals in carbon dioxide by machine learning paradigms

The CO2-based enhanced oil recovery operations not only increase the recovery factor but can result in reducing the amount of CO2 released into the atmosphere. It is possible to improve the effectiveness of CO2 injection into the reservoir by incorporating gas-soluble chemicals. Since the experimental measurement of the dissolution pressure of chemicals in CO2 [i.e., cloud point pressure (CPP)] is costly and time-consuming, it is necessary to develop a model for its reliable estimation. Although some thermodynamic-based correlations are proposed to estimate the cloud point pressure (CPP), their inaccuracy and limited ranges of application are challenging. Consequently, this study utilizes artificial neural networks, least-squares support vector regression, and adaptive neuro-fuzzy inference systems to accurately anticipate the CPP value as a function of chemical type, chemical concentration, and temperature. Both the correlation matrix analysis and standardized coefficient proved that the CPP increases by all these explanatory features so that the chemical concentration is the most important variable. Tuning the models' hyperparameters and monitoring their accuracy in the cross-validation and testing stages confirm that the general regression neural network is the most accurate paradigm for estimating the CPP value. This model anticipates 381 literature records with a mean absolute error of 8.09, an absolute average relative deviation percent of 3.27%, and a regression coefficient of 0.974 21. This straightforward machine learning paradigm not only broadens our understanding of the chemical dissolution in CO2 but also helps appropriately design the CO2-based enhanced oil recovery scenarios.