PREDICTION OF THE GLASS TRANSITION TEMPERATURES FOR POLYMERS WITH ARTIFICIAL NEURAL NETWORK

The glass transition temperature (T-g) values of three classes of vinyl polymers, i.e. polystyrenes, polyacrylates, and polymethacrylates, were predicted by using a quantitative structure-property relationship (QSPR) model constructed by back-propagation (BP) neural network. The four descriptors (the rigidness descriptor RHR resulted by hydrogen-bonding moieties group and/or rings, the chain mobility n, the molecular average polarizability alpha, the net charge of the most negative atom q(-)) were obtained directly from the polymers' monomer structures. Stepwise multiple linear regression analysis (MLRA) and artificial neural network (ANN) were used to generate the model. Simulated with the final optimum BP neural network [4-2-1], the results showed that the predicted T-g values were in good agreement with the experimental data, with a training set root-mean-square (rms) error of 20.478K (R = 0.955) and a prediction set rms error of 20.174K (R = 0.955).