Determination of binary diffusion coefficients of hydrocarbon mixtures using MLP and ANFIS networks based on QSPR method

Abstract

In this article, at first, a quantitative structure–property relationship (QSPR) model for estimation of limiting diffusion coefficients of hydrocarbon liquids, DAB0, is developed. Particle swarm optimization variable selection (PSOvs) method, together with multi-linear regression (MLR) model, selects suitable descriptors from a pool of 1124 predefined descriptors. The MLR model is trained based on a data bank of 345 experimental binary diffusivities of hydrocarbons in infinite dilution solutions. The objective function of the optimizer for descriptor selection is the multi-variant RQK function. In addition, an artificial neural network (ANN) and an adaptive neuro-fuzzy inference system (ANFIS) are developed as two nonlinear black-box estimators of DAB0 based on the descriptors obtained from PSOvs algorithm. The estimation results of these three QSPR models (PSO-MLR, ANN and ANFIS) are compared with five semi-empirical correlations. Furthermore, using 402 experimental binary diffusion coefficient data at different concentrations and temperatures, another multi-layer perceptron network (MLPN) is constructed to predict the diffusion coefficients of binary hydrocarbon liquid systems at various concentrations and temperatures. The input variables of this network are DAB0, DBA0, and mole fraction of the component A. Average absolute relative deviations (AARD) of the developed MLP-QSPR model for training, test, and whole data are respectively, 5.86%, 7.79%, and 6.32%. Indeed this model has less than 5% AARD for more than 50% of the data points. Also, the results of the MLP model designed to estimate the binary diffusivities of concentrated hydrocarbon mixtures show that the average absolute relative error for the whole of test and training data is 6.3%. The comparison of the developed MLP model with experimental data and some of semi-empirical correlations indicates that the suggested modeling scheme based on QSPR and MLP method has very good accuracies for estimation/prediction of liquid hydrocarbon diffusivity not only in infinite dilution but also at concentrated solutions.