ANN applied in the Separation of Isopropanol/Water by Azeotropic Distillation by Pressure Oscillation

Abstract

[EN] In this study, an artificial neural network (ANN) was developed to predict the concentrations of isopropanol (IPA) and diisopropyl ether (DIPE) in a pressure-swing azeotropic distillation system for the separation of isopropanol/water mixtures. To build the ANN, a simulation was validated using the open-source software DWSIM, and a sensitivity analysis was performed to determine the output variables (targets) to be predicted: IPA and DIPE molar fractions. Additionally, different experiments were carried out to obtain a set of 400 data pairs for the training and validation of the ANN. The design of the ANN was implemented in MATLAB, using the Bayesian regularization algorithm with 50 neurons in the hidden layer. The mean squared error obtained in the testing phase was 0.00186, and the regression coefficient was 0.964. To validate the ANN, an analysis of variance (ANOVA) was performed with a set of 25 data points, considering the input variables used in the ANN design. After the analysis of variance, it was concluded that the results predicted by the ANN did not present significant differences from the experimental values, with a reliability of 95%. Therefore, the developed ANN can be reliably used to predict the concentrations of IPA and DIPE in an isopropanol/water mixture separation process.