The complexity of this problem dissuades the use of a purely first-principles approach. Consequently, a hybrid approach based on quantitative structure-property relationships will be employed, using fundamental knowledge in tandem with artificial intelligence tools like computational neural networks (CNN). We will present results for the performance of the CNN for a number of bisphenol A-polycarbonate derivatives using a variety of descriptors. Topological descriptors that abstract information at the level of molecular structure have been studied using several neural network architectures such as the radial-basis function (RBF) and multilayer feed-forward. A bootstrap cross-validation procedure was employed, which removes the bias in the selection of training and test sets, minimizes the possibility of chance correlations and provides more robust generalization capabilities. About 200 learners were separately trained-and-tested using bootstrap making this approach more rigorous than a single random train-and-test partition. The estimated error for the aggregate learner is taken as the average of the error over the individual learners. Sensitivities of the networks were explored with respect to changes in the activation function, number of hidden layer neurons and the initial weight vector. For the RBF network, the glass transition temperatures of 64% of the polymers were predicted to within 10°C of their actual values (i.e. R2=0.9275 and RMS=11.6°C) while the number improved to 73% with the feedforward ensemble (i.e. R2=0.9548 and RMS=9.4°C). The results are promising despite the use of topological indices that have only a limited physical basis as descriptors for Tg. We will discuss the opportunities of using more physically relevant descriptors like flexibility, configurational entropy and packing in conjunction with CNNs for predicting Tg.