Generalized Model for Predicting Gas Permeability of Glassy Polymers and Residual Neural Networks as a Tool for Its Improvement

Abstract

The study demonstrates new opportunities for improving the prediction of gas transport properties of glassy polymers based on their chemical structure using the Database of the Topchiev Institute of Petrochemical Synthesis, Russian Academy of Sciences. A generalized linear model has been developed to predict permeability coefficients for any gas-polymer system based on structural descriptors of the polymer and gas properties, such as tabulated effective kinetic diameters of gas molecules and effective Lennard–Jones potential parameters. This model significantly expands the dataset available for predictions and the application of modern machine learning methods. The feasibility of using small residual neural networks to enhance the accuracy of linear model predictions is shown, and training such neural networks does not require significant computational resources.