Neural network based quantitative structural property relations (QSPRs) for predicting boiling points of aliphatic hydrocarbons

Quantitative structural property relations (QSPRs) for boiling points of al iphatic hydrocarbons were derived using a back-propagation neural network a nd a modified Fuzzy ARTMAP architecture. With the backpropagation model, th e selected molecular descriptors were capable of distinguishing between dia stereomers. The QSPRs were obtained from four valance molecular connectivit y indices ((1)chi(v),(2)chi(v),(3)chi(v),(4)chi(v)), a second-order Kappa s hape index ((2)kappa), dipole moment, and molecular weight. The inclusion o f dipole moment proved to be particularly useful for distinguishing between cis and trans isomers. A back-propagation 7-4-1 architecture predicted boi ling points for the test, validation, and overall data sets of alkanes with average absolute errors of 0.37% (1.65 K), 0.42% (1.73 K), and 0.37% (1.54 K), respectively. The error for the test and overall data sets decreased t o 0.19% (0.81 K) and 0.31% (1.30 K), respectively, using the modified Fuzzy ARTMAP network. A back-propagation alkene model, with a 7-10-1 architectur e, yielded predictions with average absolute errors for the test, validatio n, and overall data sets, of 1.96% (6.79 K), 1.83% (6.45 K), and 1.25% (4.4 2 K), respectively. Fuzzy ARTMAP reduced the errors for the test and overal l data sets to 0.19% (0.73 K) and 0.25% (0.95 K), respectively. The back-pr opagation composite model for aliphatic hydrocarbons; with a 7-9-1 architec ture, yielded boiling points with average absolute errors for the test, val idation, and overall set of 1.74% (6.09 K), 1.25% (4.68 K), and 1.37% (4.85 K), respectively. The error for the test and overall data sets using the F uzzy ARTMAP composite model decreased to 0.84% (1.15 K) and 0.35% (1.35 K), respectively. Performance of the QSPRs, developed from a simple set of mol ecular descriptors, displayed accuracy well within the range of expected ex perimental errors and of better accuracy than other regression analysis and neural network-based boiling points QSPRs previously reported in the liter ature.