Comparison of Neural Network and Empirical Models for Prediction of Second Virial Coefficients for Gases

Abstract

In the paper, two tools are used for estimation of second virial coefficient for gases and the obtained results are compared with the experimental data. The first tool is the computer program for estimation of second and third virial coefficients for gases and gas mixtures from basic properties of components. The computer program incorporates two empirical methods, the Tsonopoulos method to estimate second virial coefficients for nonpolar to polar pure gases and gas mixtures, and the method of Orbey and Vera to estimate third virial coefficients for nonpolar pure gases and gas mixtures. The second tool is the artificial neural network model (ANN) for correlation and prediction of second virial coefficients for gases. The neural network model was developed with the training variables: critical temperature, critical pressure, critical volume, acentric factor, dipole moment and temperature with the learning method back propagation of errors according to the best prediction error. The target variable was the second virial coefficient of gas. The neural network model has architecture (6,10,4,1). The training error was 0.3%. The network predicts the second virial coefficient with the average prediction error of 1.3%. The second virial coefficients for twenty gases are estimated with both tools. The comparison of results with experimental data shows that the computer program based on empirical methods, and the neural network model are appropriate tools for second virial coefficient prediction for gases, but more accurate results are obtained with the neural network model which gives good predictions of second virial coefficients for every gas.