Neural network protocol to predict interfacial tension for CO2/CH4/Water-Brine ternary systems under reservoir temperature and pressure ranges

The increasing importance of CO2 in reservoir engineering demands more accurate models for predicting interfacial tension (IFT). We investigated the use of neural networks to predict the IFT of ternary systems CO2/CH4/Water-Brine in a wide range of reservoir conditions, since the existing correlations do not capture all of the system's mechanisms. We introduced an optimization associated with k-fold cross validation to find optimal architecture and ensure an unbiased model. The average absolute relative error obtained is 1.33% (water) and 1.99% (brine) considerably more accurate than the best empirical models tested: 9.87% (water) and 19.11% (brine). Molecular dynamics simulation performed confirms the high superficial activity of CO2.