Predição de propriedades termodinâmicas de equilíbrio líquido-vapor com uso de redes neurais

Abstract

This work aims to evaluate the use of artificial neural networks (ANNs) in the prediction of thermodynamic vapor-liquid equilibrium (VLE) data. The use of ANNs for the prediction of thermodynamic VLE data for the chloroform-benzene-1-butanol ternary system was discussed and evaluated. For the evaluation of the efficiency, a total of 10 sets were used that mention the various conditions of prediction for ANNs in relation to thermodynamic data. These are formed by data calculated by the NRTL method, with the development of an iterative algorithm, and experimental data via literature. The data were for a system in equilibrium at pressures of 105, 205 and 303 kPa. ANNs were developed in the MATLAB® environment. A study of the best topology for the ANNs representing each of the sets was carried out. In the analysis of the results presented by ANNs, the efficiency in predicting data after training with a fully experimental set was found to result in deviations of the order of 10−2. Under conditions of use of calculated data a greater dispersion of the data was observed in relation to the expected values. Thus, it was possible to conclude that ANNs can efficiently predict thermodynamic properties of equilibrium systems, whenever trained with reliable data, presenting as an alternative for predicting data from systems that have little or no reference in the literature.