Framework Flexibility-Driven H ₂ O Diffusion in MOF-303; Mechanistic Insights from Machine-Learned Interatomic Potentials

A metal organic framework, namely MOF-303, has emerged as a promising sorbent for water harvesting applications; however, the intrinsic diffusion mechanism of adsorbed H2O molecules, crucial for reticular design of such materials, is still unknown. The computational prediction of thermodynamic properties of H2O in MOF-303 has become a common practice, but most existing literature neglects the dynamic behavior in the flexible framework. Moreover, the limitations of classical force fields fail in accurately describing vibrational states, imposing a major bottleneck in achieving realistic dynamics. In this study, we present a methodology for achieving a chemically accurate diffusion mechanism in a fully flexible framework. The methodology employs highly accurate and efficiently trained deep machine learned interatomic potentials (DP-MLIP) and molecular dynamics simulations. Accurate diffusion coefficients for H2O in flexible MOF-303 are obtained with the DP-MLIP, and the associated dynamic behavior is benchmarked against the density functional theory method (DFT) and recently reported generic foundation model (MACE) in nudged elastic band (NEB) calculations. Analysis shows close approximations in diffusion barriers predicted by DP-MLIP, DFT, and MACE, indicating these approaches capture the same sequence of adsorption-state stability. In contrast, the diffusion barrier obtained from the classical force field deviates substantially from the reference DFT. H2O molecules exhibit a higher self-diffusion coefficient (∼10×) in MD with the DP-MLIP model than the classical potential. The atomistic simulations with our trained MLIP enable a more accurate and efficient computational evaluation of the dynamic behavior of H2O molecules in flexible MOF-303, offering a new opportunity for rational design of materials with comparable flexibility.