Prediction of the Liquid-Crystalline Property Using Different Classification Methods

Abstract

This article reports a new approach to predict the liquid crystalline behavior using three classes of machine learning algorithm: eager learners (C4.5, random tree, random forest, reduced error pruning tree), lazy learners (k-nearest neighbor, non-nested generalized exemplars with prototypes), and neural networks. The performance analysis for these algorithms supposes calculating the errors both for the training and validation data, in order to determine the accuracy in building the model and its generalization capability. A large database containing 390 compounds with different units connected to the bis- and tris-phenyl aromatic, azo-aromatic, azomethinic types was used as example. The prediction of liquid crystalline property is correlated with chemical structure, type of compound, molecular weight, and a series of structural characteristics estimated by mechanical molecular simulation. Advantages and disadvantages are discussed for each method; the best results are obtained with lazy classification algorithms, especially with our original one, non-nested generalized exemplars with prototypes (NNGEP). [Supplementary materials are available for this article. Go to the publisher's online edition of Molecular Crystals and Liquid Crystals for the following free supplemental resource: Data Base]