The development of a quantitative structure property relationship (QSPR) for the prediction of dielectric constants using neural networks Original Research Article

The use of quantitative structure property relationships (QSPRs) is proposed for the calculation of dielectric constants. A data set of 497 compounds with a wide variety of functional groups is assembled. These compounds span the dielectric constant range 1–40. A total of 65 molecular descriptors is calculated for these compounds. These descriptors include the dipole moment, polarizability, counts of elemental types, an indicator of hydrogen bonding capability, charged partial surface area descriptors, and molecular connectivity descriptors. Subsets of these descriptors are used to build models in an attempt to find the best possible correlation between chemical structure and dielectric constant. A total of 70 000 models are examined. Neural networks using the Broyden–Fletcher–Goldfarb–Shanno (BFGS) training algorithm are employed to build the models. A total of 191 models have test set errors less than 2.0 and training set errors less than 3.0, where the errors are calculated as the mean of the absolute values of the residuals for sets of 97 and 350 compounds, respectively.