QSPR estimation models of normal boiling point and relative liquid density of pure hydrocarbons using MLR and MLP-ANN methods

Abstract

This work aimed to predict the normal boiling point temperature (Tb) and relative liquid density (d20) of petroleum fractions and pure hydrocarbons, through a multi-layer perceptron artificial neural network (MLP-ANN) based on the molecular descriptors. A set of 223 and 222 diverse data points for Tb and d20 were respectively used to build two quantitative structure property relationships-artificial neural network (QSPR-ANN) models. For each model, the total database was divided respectively into two subsets: 80% for the training set and 20% for the test set. A total of 1666 descriptors were calculated, and the statistical reduction methodology, based on the Multiple Linear Regression (MLR) method, has been adopted. The Quasi-Newton back propagation (BFGS) algorithm was applied in order to train the ANN. A comparison was made between the outcomes of obtained QSPR-ANN models and other well-known correlations for each property. The two best QSPR-ANN models result showed a good accuracy confirmed by the high determination coefficient (R2) values and the low mean absolute percentage error (MAPE) values ranging from 0.9999 to 0.9931 and from 0.5797 to 0.2600%, respectively for both best models (Tb and d20 models). Furthermore, the comparison between our models and the other quantitative structure property relationships (QSPR) models shows that the QSPR-ANN models provided better results. This computational approach can be applied in the petroleum engineering for an accurate determination of Tb and d20 of pure hydrocarbons.