Improved artificial neural networks (ANNs) for predicting the gas separation performance of polyimides

This study aimed to establish a quantitative structure–property relationship (QSPR) model for predicting the gas separation performance of polyimide membranes using neural networks combined with the repeat unit structure of materials. Using a data bank based on 125 polyimides, we calculated a total of 20 descriptors for all polyimides using Yampolskii's group contribution method, which divides polyimides' repeat units into their smallest groups. The number of groups contained in each polyimide is taken as the network input, and the gas permeability as the network output. Two neural network models, back-propagation (BP) and genetic algorithm-optimized back-propagation (GABP) algorithms, were used as the prediction model, and the prediction results were compared. When compared with the previous models used to predict the gas separation performance for all polymers and other machine learning (ML) models, the prediction results obtained using the GABP model are encouraging, showing a root mean squared error (RMSE) of 0.44 for CO₂, indicating that the model is applicable to polyimide. In addition, the GABP model is easy to operate, requires few parameters, it is also applicable to copolyimides. The GABP model based on the group contribution method can thus satisfactorily predict polyimides' gas separation. This is expected to be used to guide the synthesis and structure screening of polyimides for saving resources and commercialization.