This research introduces a multi-tiered multi-task learning framework for predicting gas permeability in polymers. The framework combines experimental data with simulation data using advanced polymer fingerprinting and data fusion techniques, improving the generalizability of predictive models across novel chemical spaces. The multi-task approach leverages the relationship between permeability, diffusivity, and solubility, enhancing predictive accuracy compared to traditional methods. High-throughput classical simulations complement experimental data, making the approach particularly useful when experimental data is scarce.

Brandon K. Phan, Kuan-Hsuan Shen, Rishi Gurnani, Huan Tran, Ryan Lively, Rampi Ramprasad