Predicting and considering properties of general polymers using incomplete dataset

Abstract

Polymer properties are usually more difficult to predict than those of small molecules due to them forming superstructures. In this work, we aimed at finding a versatile approach to predict multiple polymer properties using imperfect data with missing values. The dataset was hierarchically clustered on the basis of two independent factors: polymer properties and polymer structures. In polymer property-based clustering, visualizing relations of polymers was found to be an effective way of estimating the difficulty of polymer property prediction. In polymer structure-based clustering, each cluster could be formed based on the structural features. Thus, the clustering contributed to understanding structural characteristics of monomer unit structures. In addition to analyzing the data set in an unsupervised manner, we constructed polymer properties prediction models based solely on the information of monomer unit structures. Partial least squared (PLS) regression models could predict density, glass transition temperature and dissolution parameter with high accuracy. We also propose approach to evaluate obtained model using data already prepared.