Abstract
In the development of new polymers, it is important to predict properties of polymers.But there are problems in existing methods to predict polymer properties. Molecular dynamics simulation takes long time and gives systematic error, and group contribution method requires endless correction terms to get exact values. Neural network has been applied to the prediction of glass transition temperatures of polymers. Network was trained by entering numbers of 12 kinds of bonds in monomer unit structures as input descriptors, and a leave-one-out test revealed that the network can well predict glass transition temperatures ranging above 600K with an average error of 29K.
