Molecular structure incorporated deep learning approach for the accurate interfacial tension predictions

Characterization of the interface of a two-phase system by interfacial tension (IFT) imposes a great impact on the chemical and environmental engineering. In this work, a deep neural network (DNN) approach was developed to estimate IFT of water-hydrocarbon and water-alcohol interfaces. The predictive power of this approach for IFT was found to be much more improved than those of the previously proposed empirical correlations, both qualitatively and quantitatively. The input vector of two-phase systems generally contains five parameters, including critical temperature, critical pressure, and density difference, in addition to temperature and pressure. In this approach, a line notation describing the molecular structure of chemical species was also taken as an input. The most accurate results with the root-mean-square error (RMSE) of 1.28 mN/m are acquired as all six parameters are included. However, our analyses show that density difference and molecular structure are much more important than the critical properties. As a result, the DNN approach with the input vector involving molecular structure, temperature, and pressure only is able to yield sufficiently accurate results (RMSE 1.71 mN/m), and can successfully depict the descending, ascending, and concave dependences of IFT on temperature.