Estimation of CO2–brine interfacial tension using an artificial neural network

Abstract

Experimental determination of CO2–brine interfacial tension (IFT) usually requires expensive apparatus and sophisticated interpretation procedure and is time-consuming. Hence, it is of practical importance to develop an accurate and reliable model for determining the CO2–brine IFT. This paper presents the use of feed forward artificial neural network (ANN) to accurately estimate CO2–brine IFT based on a database acquired from previous literature. The database consists of a total of 1716 CO2–brine IFT datasets that cover relatively large ranges of pressure (0.1–60.05MPa), temperature (5.25–175°C), total salinity (0–5molkg−1) and mole fractions (0–80%) of impure components. Six independent variables were considered to develop the IFT estimation model: pressure, temperature, monovalent cation (Na+ and K+) molality, bivalent cation (Ca2+ and Mg2+) molality in brine, and mole fractions of N2 and CH4 in injected CO2 streams. The ANN topology was optimized by trial-and-error in order to enhance its capability of generalization and the optimal one was determined to be 6-10-20-1 (10 and 20 neurons in the first and second hidden layers, respectively). The accuracy of the proposed ANN model was highlighted by four evaluation matrices, namely mean absolute error (MAE), mean absolute relative error (MARE), mean squared error (MSE), and determination coefficient (R2) between the measured and estimated IFT. The ANN model was further compared against four empirical IFT correlations developed in previous studies. It was observed that the ANN model outperforms significantly the empirical correlations and provides the most accurate IFT reproduction with respect to pure CO2–pure water, pure CO2–brine and impure CO2 systems.