A Neural Network QSPR Model for Accurate Prediction of Flash Point of Pure Hydrocarbons.

Abstract

The present study introduces a QSPR model to predict the flash point of pure organic compounds from diverse chemical families. We used the Maximum-Relevance Minimum-Redundancy (MRMR) as an efficient descriptor selection algorithm to select 20 the most effective out of 1926 calculated descriptors. The selected descriptors and their combination with the normal boiling point data were used as model inputs and their correlation with FP was mapped using feedforward artificial neural networks. Studying various models, the best result was obtained by a neural network with 2 neurons in the hidden layer for which a combination of the selected descriptors and normal boiling point data were used as model inputs. Evaluating the performance of this model for a dataset of 727 compounds resulted in average absolute relative errors of of 1.36 %, 1.34 %, 1.44 % and 1.42 % and average absolute deviations of 4.48, 4.41, 4.75 and 4.66 K for the overall, training, validation, and test datasets, respectively.