Prediction of gas permeation through polymers with intrinsic porosity using a hybrid neutral network-particle swarm model

This study develops a quantitative structure-property relationship (QSPR) model using a hybrid neural network and particle swarm optimization (PSO) to predict the gas separation performance of 120 polymers of intrinsic microporosity (PIMs). Over 5000 descriptors, including topological, constitutional, functional groups, and geometrical properties, were computed using alvaDesc software. Genetic algorithm optimization combined with partial least squares regression was used to select relevant descriptors for predicting PIM permeability to N2, CH4, and CO2. A hybrid neural network model with particle swarm optimization-based backpropagation (PSO-BP) algorithms was used for permeability prediction, and the results were compared to experimental published data. The PSO-BP model showed promising results, with root mean squared error (RMSE) values of 0.0048, 0.000743, and 0.0045 for CO2, N2, and CH4, permeabilities respectively. Key descriptors for predicting PIM permeability are associated with multiple physicochemical properties, including GATS, 3D Morse, TDB, SpMax, MATS, CATS3D, RDF, and ATS descriptors. CO2 permeability prediction requires more 3D descriptors than N2 and CH4.