(572d) A Combined Deep Learning and Classical Potential Approach for Modeling Diffusion in Uio-66

We have developed an accurate deep learning potential (DP) for UiO-66, a metal-organic framework, that is trained using data generated from density functional theory (DFT) calculations to reproduce energies and forces on atoms with close to DFT accuracy. The generation of training data was split into two steps. First, we employed an iterative data generation technique that explores compressed and expanded configurations of UiO-66 based on the DP's ability to accurately reproduce the DFT-computed equation of state. This initial step was performed to ensure accurate description of mechanical properties of UiO-66. Second, we focused on improving the description of framework flexibility by generating data from high temperature active learning iterations assisted by DFT-molecular dynamics (MD) simulations. The resultant DP achieved very low root-mean squared errors in predicting energies 1.65 10^-3 eV/atom and atomic forces (7.8 10^-2 eV/Å). The DP has a 0.012 % error in predicting the lattice constant and 0.1 % error in predicting the bulk modulus compared to DFT. We developed a hybrid DP and Lennard-Jones (DP/LJ) potential to account for adsorbent-adsorbate interactions and used this to compute diffusivities of guest fluids like Ne and Xe through UiO-66. This hybrid approach does not require refitting the DP for new adsorbates. Self-diffusion coefficients were calculated at finite Ne-loading from the DP/LJ potential and from DFT-MD. The DP/LJ results are in excellent agreement with DFT-MD simulations. Diffusion coefficients were also computed using two classical force fields for UiO-66 and were found to give similar results. Zero-loading simulations were performed using DP/LJ at three different temperatures in order to estimate the Ne diffusion barrier. The DP/LJ calculations give an activation energy for Ne diffusion of 2.67 0.18 kJ/mol. The DP/LJ was also used to perform zero-loading diffusion simulations of larger guest fluids like Xe into the MOF to demonstrate the DP’s ability to model the framework flexibility of UiO-66. The self-diffusion coefficient of Xe predicted by DP/LJ was compared to classical force fields. Our hybrid DP-classical potential approach can be applied to other MOFs and other adsorbates, making it possible to use an accurate DP generated from DFT simulations of an empty MOF in concert with existing classical potentials for adsorbates to model adsorption and diffusion within the MOF. Our results show how very accurate potentials for porous adsorbents can be generated and then applied to modeling diffusion of various adsorbates without generating potentials for each adsorbent-adsorbate pair.