Polymer Property Prediction and Optimization Using Neural Networks

N.K. Roy, W.D. Potter, D.P. Landau
2006 IEEE Transactions on Neural Networks  
Prediction and optimization of polymer properties is a complex and highly non-linear problem with no easy method to accurately predict polymer properties directly and accurately. The problem is even more complicated in high molecular weight polymers like engineering plastics which have the greatest use in industry. The effects of modifying a mer (polymer repeat unit) on the polymerization and the resulting polymer properties is not as easy a problem to investigate experimentally given the large
more » ... number of possibilities. This severely curtails the design of new polymers with specific end use properties. Another aspect in the development of useful materials is the use of polymer blending. Here again predicting miscibility or compatibility together with resulting blend properties is difficult, but imperative in order to get a useful product. Since miscibility can be controlled using low molecular weight compatibilizers the problem for experimentalists to find the correct working combination is further exacerbated. In this paper we show how properties of modified mers can be predicted using Neural Networks. In an earlier paper[1] we have given an overall approach to polymer blend design in which the use of Neural Networks was but one stage in the process that also made use of genetic algorithms and Markov chains to accurately predict blend miscibility. Here we present in great depth the first step which is the prediction and optimization of modified polymer properties using Neural Networks. We utilize a large database of polymer properties and employ a wide variety of networks ranging from backpropagation networks to unsupervised self-associating types. Based on extensive training and comparisons we are able to select particular networks that can be used to predict accurately specific polymer properties. We are able to classify the networks in the design into groups that range from those that provide quick training to those that provide excellent generalization. We are also able to show how the extensive databases available on polymers can be easily used to accurately predict and optimize polymer properties.
doi:10.1109/tnn.2006.875981 pmid:16856662 fatcat:ubgljvsp7jen5n66h4ttqgiveu