Machine learning assisted modeling of interfacial tension in the system N2/Brine

Abstract

The CO2 rising in the atmosphere causes serious greenhouse effects on severe and sustainable global warming. Carbon capture is one of the most viable approaches accepted to achieve sustainable development goals. On the other hand, flue gases contain large amounts of CO2 and N2. For the purpose of CO2 sequestration in depleted reservoirs or saline aquifers, a deep understanding about the interaction between these components with brine is vital. Among the most important parameters, which can be investigated for such applications, is interfacial tension (IFT) in these systems. In the current study, we have measured IFT between N2 and different brines including monovalent and divalent ions at various temperatures and pressures at different salinity ranges. Various brines containing different salts such as NaCl, Na2SO4, KCl, MgCl2, and CaCl2 were synthesized in the range of salinity between 10000 ppm and 50000 ppm. The temperature and pressure ranges of our study were 30–70 °C, and 0.1–30 Mpa, respectively. IFT700 (Vinci, France) instrument was used for our experiments. The obtained experimental data were then analyzed using different machine learning algorithms. The prediction of IFT as a response variable (in the range of 31.39–74 mN/m) with independent variables including temperature (K), pressure (Mpa), and molar concentrations of different anions and cations (M) was carried out using adaptive neuro-fuzzy inference system (ANFIS) and artificial neural network (ANN) with feed-forward architecture models which was trained using particle swarm optimization (PSO) algorithm as an evolutionary algorithm. Based on the present study results, the IFT was predicted accurately by ANFIS + PSO and ANN + PSO endorsed with good performance parameters. The coefficient of determination (R2), mean squared error (MSE), mean absolute error (MAE), and average absolute relative deviation (AARD%) were calculated to assess the accuracy of the models. Both models showed very good performance, with R2 coefficient 0.9898 and 0.9938 for test data of ANN + PSO and ANFIS + PSO, respectively. As it can be seen from the results, ANFIS and ANN methods were very successful for predicting the behavior of the complex systems studied in this work, while the ANFIS + PSO model had a better predictive power. It can be explained by the fact that ANFIS is a modeling technique that combines the power of artificial neural network with fuzzy logic theory techniques. It can apply FIS to define hidden layers and improve its predictive ability.