Neural networks for the prediction of polymer permeability to gases

Abstract The objective of this study was to develop a quantitative structure-property relationship (QSPR) using an artificial neural network (ANN) in order to improve the prediction of gas permeability coefficients of membranes for some major industrial gases, such as O 2 , N 2 , CO 2 , and CH 4 . Using a polymer data bank based on 149 polymers, a total of 21 descriptors were calculated for all polymers using the group contribution technique of Yampolskii, which decomposes polymers’ structures into their smallest groups and characterizes a polymer's repetitive units. These molecular descriptors represent the network inputs, whereas the gas permeability coefficients of the polymers constitute its output. Permeability prediction results of various calculations obtained with the developed ANN models provided highly encouraging results. Correlation factors R of 0.999, 0.999, 0.984 and 0.999 and low root mean square errors (RMS) of 1.054, 2.635, 150, and 2.46 were reached for N 2 , O 2 , CO 2 and CH 4 , respectively. Significant improvements in the predictions of the proposed ANN models were observed compared with the previously published results, expressed in terms of performance factor C, which generally ranged from 0.0075 to 0.06 for all studied gases. A best value of 0.77 was achieved for the CO 2 permeability prediction by a published study in this field.