Neural Network Model of Surface Tension of Hydrocarbon–Alcohol Binary Mixtures

Abstract

Since the correlation of surface tension of hydrocarbons and alcohol mixtures has great importance, this study is aimed to propose an artificial neural network model to correlate this thermodynamic property. To calculate the best network architecture and the optimal number of neurons, five sets of input variables and two transfer functions are examined. The results reveal that a feed-forward network with the structure of 5-14-1 and logsis and purelin as transfer functions for hidden and output layers, respectively, leads to the best accuracy. Moreover, it is revealed that choosing temperature, mole fraction, molecular weight of hydrocarbons, molecular weight of alcohols, and critical temperature as input variables can be efficient for an accurate correlation of the surface tension of selected chemicals. The results of the proposed artificial neural network model are also compared to the Shereshefsky and Langmuir thermodynamic models. The results obtained from 34 binary mixtures show the generality and acceptable accuracy of the proposed feed-forward network (with an average absolute relative deviation [AARD] of 0.36%) compared to the Shereshefsky (with an AARD of 0.37%, obtained for 31 binary mixtures) and Langmuir (with an AARD of 0.52%) thermodynamic models.