Machine Learning Models for Predicting Molecular Diffusion in Metal-Organic Frameworks Accounting for the Impact of Framework Flexibility

Yuhan Yang; Zhenzi Yu; David S Sholl

doi:10.1021/acs.chemmater.3c02321

Machine Learning Models for Predicting Molecular Diffusion in Metal-Organic Frameworks Accounting for the Impact of Framework Flexibility

Chem Mater. 2023 Nov 22;35(23):10156-10168. doi: 10.1021/acs.chemmater.3c02321. eCollection 2023 Dec 12.

Authors

Yuhan Yang^{1

2}, Zhenzi Yu¹, David S Sholl^{1

3}

Affiliations

¹ School of Chemical & Biomolecular Engineering, Georgia Institute of Technology, Atlanta, Georgia 30332-0100, United States.
² School of Chemical Engineering and Technology, Hainan University, Haikou 570228, China.
³ Oak Ridge National Laboratory, Oak Ridge, Tennessee 37831, United States.

Abstract

Molecular diffusion in MOFs plays an important role in determining whether equilibrium can be reached in adsorption-based chemical separations and is a key driving force in membrane-based separations. Molecular dynamics (MD) simulations have shown that in some cases inclusion of framework flexibility in MOF changes predicted molecular diffusivities by orders of magnitude relative to more efficient MD simulations using rigid structures. Despite this, all previous efforts to predict molecular diffusion in MOFs in a high-throughput way have relied on MD data from rigid structures. We use a diverse data set of MD simulations in flexible and rigid MOFs to develop a classification model that reliably predicts whether framework flexibility has a strong impact on molecular diffusion in a given MOF/molecule pair. We then combine this approach with previous high-throughput MD simulations to develop a reliable model that efficiently predicts molecular diffusivities in cases in which framework flexibility can be neglected. The use of this approach is illustrated by making predictions of molecular diffusivities in ∼70,000 MOF/molecule pairs for molecules relevant to gas separations.