Since its inception, the non-equilibrium lattice fluid (NELF) model has become a vital tool in correlating and predicting the gas solubility behaviour in glassy polymeric membranes. But like its equilibrium variant, the NELF model is highly constrained by the availability of the pure polymer characteristic parameters, which are not always convenient to obtain as the need arises. In this study, we provide a proof-of-concept for building a machine learning-based group contribution method (ML-GC) for the Sanchez–Lacombe equation of state (EoS) pure polymer parameters. The ML-GC model was built using a modified version of the Marrero and Gani’s method, which incorporates machine learning regression into the GC parameterisation process. The final model was capable of reproducing the parameters of a randomly selected test set of polymers, with a diverse range of chemical structures. The resultant average AARD% of the predicted densities in this set is 5.59%, with no polymer exceeding 15%. Moreover, to test the model’s capabilities in estimating the parameters of high glass transition temperature polymers, we predicted a priori the characteristic parameters of 6 polyimides from the knowledge of their molecular structure. The ML-GC parameters were also incorporated into the NELF model to predict the infinite dilution solubility coefficients (S0) of some of these polymers and the results were validated against experimental data. Furthermore, the ML-GC-NELF model was also used for the first time to represent effectively the gas solubility isotherms in PIM-PI-SBI and PIM-PI-EA with relatively small magnitudes of the binary interaction parameters (kij). Despite the small data-set used herein, the model performance was satisfactory, however, as more data are being published in literature, the proposed ML-GC model has the potential of providing even more accurate predictions for a wider range of polymers, ultimately leading to lesser reliance on experimental data for modelling the gas sorption.
Read full abstract