Bg. Sumpter et Dw. Noid, NEURAL NETWORKS AND GRAPH-THEORY AS COMPUTATIONAL TOOLS FOR PREDICTING POLYMER PROPERTIES, Macromolecular theory and simulations, 3(2), 1994, pp. 363-378
A new computational methodology is presented for making rapid and accu
rate predictions of chemical, physical and mechanical properties of po
lymers based on their molecular structure. The method uses a set of to
pological indices derived from graph theory to numerically describe th
e structure of a monomeric repeating unit for a given polymer (structu
ral descriptors) and relates these indices to a set of polymer propert
ies by utilizing an artificial neural network. The neural network is a
ble to efficiently formulate all of the correlations (i. e., between s
tructural descriptor-property, property-property, structural descripto
r-structural descriptor: both linear and nonlinear dependencies) neces
sary to make accurate predictions. Results have been obtained for up t
o 9 properties of 357 different polymers with an average prediction er
ror of < 3% and a maximum error of 12%, demonstrating superiority over
other quantitative structure/property relationships for polymers.