Improved quantitative structure property relationships for the prediction of dielectric constants for a set of diverse compounds by subsetting of the data set

In a recent publication we explored the development of quantitative structure property relationships for the calculation of dielectric constants, which resulted in a general model for a wide range of compounds. Our current work explores the division of the set of compounds into eight more homogeneous subsets for which local models are developed. The full data set consists of 454 compounds with dielectric constants ranging from 1 to 40. A pool of up to 16 molecular descriptors is calculated for each of the eight data sets. The descriptors include dipole moment, polarizability, counts of elemental types or functional groups, charged partial surface area, and molecular connectivity. All possible 4-16 descriptor models are calculated for each of the eight data sets, and the best models are selected and compared to the results obtained from the best general model for all 454 compounds. Neural networks using the Broyden-Fletcher-Goldfarb-Shanno training algorithm are employed to build the models. The resulting combined mean test set error for the eight local models of 1.31 is significantly better than the mean test set error of 1.85 for the general model.