A neural network approach to prediction of glass transition temperature of polymers

Polymeric materials are finding increasing application in commercial optical communication systems. Taking advantage of techniques from the field of artificial intelligence, the goal of our research is to construct systems that can computationally design polymer formulations, including polymer optical fibers, with specified desirable consumer characteristics. Through the use of an extensive structure-property correlation database, properties of polymers can be predicted by an artificial network and the structure of novel polymers with desired properties can be optimized by a genetic algorithm. In this paper, we are focusing on one of the parameters, glass transition temperature (T-g) that influences a desired outcome in polymer optical fibers. Performance of such fibers can be optimized by engineering a polymer to exhibit a lower refractive index and T-g This paper compares and discusses a neural network model and a linear model that have been developed to correlate T-g and repeating units of polymers. A neural network and multiple linear regression analysis were used in the study. A set of descriptors, chosen based on previous studies on the relations between T-g and polymer structure, were used to describe the structure of repeating units, individual bond energies, and intermolecular forces, especially hydrogen bonding, which is the strongest intermolecular force and exerts the greatest influence on T-g compared with other intermolecular interactions. A comprehensive neural network model with 28 descriptors was developed to predict T-g values of 6 randomly selected polymers from a database containing 71 polymers. The network was trained with the remaining 65 polymers and had a typical training root mean square error of 17 K (R-2 = 0.95) and prediction average error of 17 K (R-2 = 0.85). A linear regression model developed for comparison had an average error of 30 K (R-2 = 0.88). (c) 2007 Wiley Periodicals, Inc.