A GMDH-type neural network with multi-filter feature selection for the prediction of transition temperatures of bent-core liquid crystals

Abstract

A novel strategy for the prediction of the transition temperature of bent-core liquid crystals (LCs) based on the combination of multi filter feature selection and group method of data handling (GMDH) type neural networks is reported. An entire set of 243 compounds was randomly divided into a training set of 207 compounds and a test set of 36 compounds. Descriptors were selected from a pool of 2D, and two pools of 2D and 3D ones, optimized by molecular mechanics (MM) and semi-empirical (SE) method. The reduction of the pool of descriptors was performed using multi filters based on chi square and v-WSH algorithm, while the final subset selection was performed by GMDH algorithm during the learning process. The obtained 2D, MM and SE GMDH models have 11, 13 and 16 descriptors, respectively, and demonstrate good generalization and predictive ability (R² = 0.92). The final models were subjected to a randomization test for validation purpose. Those models appear to be not only suitable for prediction, but they also allow the identification of key structural features that alter the transition temperature of bent-core LCs.