Associative Neural Network and Multi Linear Regression Based Quantitative Structure Property Relationship for Modeling Viscosity of Alcohols

Abstract

A quantitative structure property relationship (QSPR) study was conducted based on molecular descriptors derived from molecular structures have been used for the prediction of viscosity of alcohols at 20 ºC. To perform this research, a set of 35 alcohols as data series was selected then variable Zagreb index (^vM₂), number of carbon atoms (N_c), molecular weight (MW) and number of hydroxyl groups (N_OH) for data series was calculated. Variable Zagreb index involving a parameter λ to be determined during regression fitting. The optimal value of the variable λ was determined by minimizing the standard error of regression. Multi linear regression (MLR) and associative neural network (ASNN) methods were used to construct the linear and non-linear prediction models respectively. The results were cross-validated by leave-one-out (LOO) procedure. The predicted results are in good agreement with the experimental viscosity of alcohols. Comparison of these two methods reveals that those model obtained by ASNN are better.