The prediction of glass transition temperature of polycarbonates using physical descriptors and neural networks

Shivani Syal, Purdue University

Abstract

The accurate prediction of polymer properties from the chemical structure of their monomeric repeat units is a challenging problem in polymer science and technology, where small changes in properties have important engineering consequences. This thesis focuses on the development of a predictive model to describe how variations in structure affect a number of important engineering properties for a series of polycarbonate resins. The initial emphasis is on glass transition temperature, an important property that dictates the processing and ultimate end use application of such polymers. Consequently, we have employed a hybrid model based on quantitative structure-property relationships, using fundamental knowledge in tandem with artificial intelligence tools like computational neural networks. Descriptors that abstract information at the level of molecular structure (primary model) have been studied using several neural network architectures such as the radial-basis function and multilayer feed-forward (secondary model). We have proposed the use of free volume and conformational entropy as the key fundamental descriptors for our primary model. The free volume was determined as the difference between the solvent-accessible and the van der Waals volume of the molecule. The solvent-accessible volume was calculated as the volume enclosed by the centre of a probe sphere, as it rolls over the van der Waals envelope of the molecules. These calculations were performed on an ensemble of 10,000 conformers generated by sampling the entire conformational space of the polymeric molecules by a Stochastic Proximity Embedding algorithm. The conformational entropy was also determined for this ensemble of low energy molecular configurations, where the probability of each conformer state was calculated as an appropriate Boltzmann distribution. A neural network model was developed to predict the glass transition temperature of the polycarbonates using bootstrap cross-validation procedure, which removes the bias in the selection of training and test sets, minimizes the possibility of chance correlations and provides more robust generalization capabilities. The estimated error for the aggregate learner was taken as the average of the error over the individual learners. The resulting predictions are within a root mean square value of 30°C. Ultimately, the true merit of such an approach lies in the rational design of new molecules by searching through a library of synthesizable monomers using an inverse solution methodology, such as genetic algorithm, to identify promising candidates possessing a desired set of properties.

Degree

Ph.D.

Advisors

Venkatasubramanian, Purdue University.

Subject Area

Chemical engineering

COinS