Polymer-based membranes are crucial for gas and solvent separation in applications such as carbon capture, water purification, and drug delivery. A key performance indicator is gas permeability (P), defined by the solution-diffusion model as the product of gas diffusivity (D) and solubility (S): P = DS. Accurately and rapidly predicting gas permeability across diverse gases and polymer chemistries is essential for materials discovery. Traditional experimental methods (constant volume permeation) are time-consuming and resource-intensive. While classical molecular dynamics (MD) simulations offer an alternative, their accuracy is limited by force field approximations and computational timescales. Machine learning (ML) methods have shown promise in predicting polymer properties but often lack generalizability when applied to new chemical spaces. This study addresses these limitations by developing a novel multi-task learning framework.

Previous ML models for gas permeability prediction in polymers primarily relied on experimental data, exhibiting robustness within known chemical domains but limited reliability when extrapolated to new spaces. Early studies used simple features like temperature and pressure, while more recent approaches incorporated detailed structural features (fingerprints) representing the polymer's chemical structure. Although these methods improved accuracy, predicting properties outside the training data remained a significant challenge. The current study builds upon these efforts by integrating both experimental and computational data, enabling more robust and generalizable predictions.

The researchers developed a multi-task (MT) learning framework that leverages data fusion techniques and combines high-fidelity experimental data with abundant low-fidelity simulation data. A high-throughput simulation pipeline was created using molecular dynamics (MD) and Monte Carlo (MC) simulations with the LAMMPS package and the GAFF2 force field to generate data for gas diffusivity (Dsim) and solubility (Ssim). Simulated permeability (Psim) was then calculated from Dsim and Ssim using the solution-diffusion model. Experimental data (Pexp, Dexp, Sexp) for six gases (CO2, CH4, O2, N2, H2, and He) across 820 polymers was curated from 84 publications. The data was then used to train polyGNN, a multitask graph neural network model that automatically generates fingerprints from SMILES strings. Four models were developed to evaluate the framework: a single-task (ST) model using only experimental permeability data, and three multi-task models (MT-1, MT-2, MT-3) integrating different combinations of experimental and simulated data for permeability, diffusivity, and solubility. Model performance was evaluated using the coefficient of determination (R²) and the order of magnitude error (OME).

The multi-task learning approach significantly outperformed the single-task model, particularly when trained on limited data. MT-1, which incorporated simulated permeability data, demonstrated improved performance over the ST model, highlighting the value of data fusion. MT-2, which included experimental data for diffusivity and solubility, showed a dramatic improvement in predictive accuracy (higher R² and lower OME), indicating the advantage of leveraging correlated properties. MT-3, incorporating all available experimental and simulated data, achieved the highest accuracy (average R² of 0.96 and average OME of 0.10). A comparison with a previous state-of-the-art model deployed at Polymer Genome showed considerable improvement in accuracy across 13 polymer classes, with R² values exceeding 0.90 for all classes in the new model. The new model also increased the number of polymers covered from 315 to 1050, and the total number of data points from 1501 to 6788. Robeson-type trade-off plots were generated for gas permeability, diffusivity, and solubility for over 13,000 known polymers, revealing potential candidates for gas separation applications and highlighting regions of high prediction uncertainty where additional data is needed.

The findings demonstrate the effectiveness of multi-task learning for predicting gas transport properties in polymers. The integration of simulation data significantly enhances the predictive capabilities of the model, particularly in data-scarce regions, addressing the challenge of extrapolating to new chemical spaces. The inclusion of correlated properties (diffusivity and solubility) further improves accuracy by leveraging underlying physical relationships. The superior performance of the MT model compared to the ST model and the previous state-of-the-art model highlights the advantages of data fusion and multi-task learning. The Robeson-type trade-off plots provide valuable insights for the design of high-performance polymer membranes, guiding the selection of polymers with desirable permeability and selectivity.

This research presents a novel multi-task learning framework that significantly improves the prediction of gas transport properties in polymers. The combined use of experimental and simulation data, along with the incorporation of correlated properties, leads to highly accurate and generalizable models. This approach is particularly valuable in scenarios with limited experimental data, accelerating the discovery and design of advanced polymer membranes. Future work could focus on expanding the dataset to include more diverse polymers and gases, improving the simulation accuracy, and exploring other relevant polymer properties.

The accuracy of the simulation data depends on the quality of the force field used. The model's predictions may be less reliable in chemical spaces underrepresented in the training data, as highlighted in the Robeson-type trade-off plots. The experimental data used was collected under various conditions, potentially introducing variability. Future refinements could focus on incorporating additional data from standardized experiments and more advanced simulation techniques.