The use of neural networks on VLE data prediction

Abstract

The neural network model is employed to predict the vapor-liquid equilibrium (VLE) data for six different binary systems having different chemical structures and solution types (azeotrope-nonazeotrope) in various conditions (isothermal or isobaric). A model based on a feed-forward back-propagation neural network is proposed. Only half of the experimentally determined VLE data are assigned to the designed framework as training patterns in order to estimate the VLE data of the whole system in given conditions. The VLE data are also calculated by the UNIFAC model, a calculation method widely used in this field. The mean deviations from the experimental data are determined for both the models. It is observed that the data found by neural network model gives an excellent agreement with the experimental data, while the UNIFAC model shows deviations, particularly at low pressures. In fact the neural network model can be treated as a potent means for VLE data prediction in a fast and reliable way, compared to the conventional thermodynamical models.