Janhavi Nistane1,Lihua Chen1,Youngjoo Lee1,Kuan-Hsuan Shen1,Ryan Lively1,Rampi Ramprasad1
Georgia Institute of Technology1
Janhavi Nistane1,Lihua Chen1,Youngjoo Lee1,Kuan-Hsuan Shen1,Ryan Lively1,Rampi Ramprasad1
Georgia Institute of Technology1
Manufacturing polymers, plastics, fibers, solvents, and fuel additives in numerous industries rely extensively on starting materials like benzene and its derivatives, including toluene, ethylbenzene, and the xylene isomers. To use petroleum-based products effectively, they often require separation into different components. This separation is achieved through energy-intensive thermal distillation processes, accounting for 10-15% of global energy usage. However, emerging membrane-based technologies have the potential to achieve these separations with about one-tenth of the current energy consumption<sup>1</sup>. A polymer membrane essentially acts as a selective filter for solvents based on their diffusion behavior. Solvents with low diffusivity are retained by the polymer membrane, while those with higher diffusivity permeate through. For a successful separation, observing a minimum two-order difference in diffusivity between the solvents as it diffuses through the polymer membrane is necessary. The main challenge associated with membrane-based separations is identifying a tailored membrane material suitable for a specific solvent-solvent separation. Given the unimaginably vast chemical polymer space, conventional experimental methods to design new polymers are slow. Hence, we employ AI-aided material design that can efficiently explore the vast chemical polymer space in a very time and cost-effective manner, as has been achieved in the design of dielectric polymers in the past.<sup> 2,3</sup> Around 10,000 potential polymer membrane candidates will be generated using our virtual forwarded synthesis, as an initial step to AI-aided material design. These hypothetical candidates will be screened using our best-in-class ML polymer-solvent diffusivity predictor, which has been trained on available experimental data, and our computationally generated data via molecular dynamics simulations. The ML polymer-solvent diffusivity predictor is trained on around 1200 data points (experimental and simulated) for 91 polymers and 50 solvents. This scheme to generate computational data is highly scalable, and we continue to keep augmenting new data to explore more polymer chemistries. Upon identifying promising candidates that meet the proposed screening criteria, experimental validation will follow. This approach will provide insights into the power, as well as the limitations of AI-driven materials design.<br/><br/>1.D. Sholl et al., Nature, vol. 532, pp. 435-437, 2016. DOI: 10.1038/532435a.<br/>2.C. Wu et al., ACS Appl. Mater. Interfaces, vol. 13, no. 45, pp. 53416-53424, 2021. DOI: 10.1021/acsami.1c11885.<br/>3.C. Wu et al., Matter, vol. 5, no. 9, pp. 2615-2623, 2023. DOI: 10.1016/j.matt.2022.06.064.