Molecular structure based property modeling: Development/ improvement of property models through a systematic property-data-model analysis

The objective of this work is to develop a method for performing property-data-model analysis so that efficient use of
knowledge of properties could be made in the development/improvement of property prediction models. The method includes: (i) analysis of property data and its consistency check; (ii) selection of the most appropriate form of the property model; (iii) selection of the data-set for performing parameter regression and uncertainty analysis; and (iv) analysis of model prediction errors to take necessary corrective steps to improve the accuracy and the reliability of property models. To make the property-data-model analysis fast and efficient, an approach based on the “molecular structure similarity criteria” to identify molecules (mono-functional, bi-functional, etc.) containing specified set of structural parameters (that is, groups) is employed. The method has been applied to a wide range of properties of pure compounds. In this work, however, the application of the method is illustrated for the property modeling of normal melting point, enthalpy of fusion, enthalpy of formation, and critical temperature. For all the properties listed above, it has been possible to achieve significant improvements in the performance of their models. The improved model for enthalpy of formation yields an average absolute deviation of 1.75 kJ/mol which is well within the required chemical accuracy. All of the available experimental data-points are used for the regression purpose. However, a method for selecting a minimum data-set for the parameter regression is also discussed for the cases where it is preferred to retain some data-points from the total data-set to test the reliability of predictions for validation purposes.