Predicting the transition temperature of smectic liquid crystalline compounds from their structure using artificial neural networks

Abstract

The derivation of material properties of chemical compounds directly from their chemical structure can be used to plan chemical syntheses more efficiently. Here, feed forward–back propagation neural networks are devised to predict the transition temperatures of smectic liquid crystalline compounds based on a set of data that contains 6304 different structural patterns. The trained networks were tested with 1575 smectic liquid crystalline compounds that the networks had not seen before. Four different network architectures were trained to predict the transition temperatures. All networks had the capability to predict a significant portion of the transition temperatures with small deviations. The network with 10 hidden neurons and one output neuron has a high recognition rate and predicts the transition temperatures of about 85% of the structures unknown to the network with an error of ⩽20 °C. In contrast, the network with 100 hidden neurons and 370 output neurons makes more precise predictions of the transition temperatures indicated by a low standard deviation of 14.3 °C and by the fact that only 8.3% of the tested structures produced an error of more than 20 °C. However, the latter network gives answers only for 79.4% of the structures in the test set.