Janhavi Nistane1,Lihua Chen1,Kuan-Hsuan Shen1,Rampi Ramprasad1
Georgia Institute of Technology1
Janhavi Nistane1,Lihua Chen1,Kuan-Hsuan Shen1,Rampi Ramprasad1
Georgia Institute of Technology1
Understanding the diffusion behavior of organic molecules through polymers is critical for many applications, including fuel production, coatings, and drug delivery systems. Experimental methods to measure diffusivity are usually time and cost-intensive. Further, available experimental data on organic molecular interactions and diffusivity is sparse and covers only limited chemical space. Here, we present a multi-task machine learning (ML) approach to instantaneously predict solvent diffusivity in polymers. The multi-task ML model is trained simultaneously on both available experimental and computational data generated by us using high-throughput classical molecular dynamics simulations. The experimental data is regarded as high-fidelity data; however, this data is only available across a limited chemical space. On the other hand, simulations offer the flexibility to obtain diffusivity values across a vast chemical space. Although there is only semi-quantitative agreement between simulations and experimental values, correlations that do exist between the two classes of data is exploited by the multi-task neural network architecture. This approach tackles the problem of data scarcity while offering a scalable scheme for further data augmentation, enabling the model to co-learn and make predictions more effectively.