Exploring foundational machine learned potentials for treating the high temperature dynamics of metal-organic frameworks

Metal-organic framework (MOF) derived materials, formed through high temperature processes, show great potential as catalysts. However, knowledge of the structure-property relationships between the initial MOF and final MOF-derived catalyst is limited, as their amorphous nature challenges standard structural characterization methods. Additionally, current molecular simulation methods, such as ab initio molecular dynamics, are computationally demanding and unable to follow pyrolysis over large temporal and spatial scales. One solution is the use of neural network approaches for learning interatomic potentials from density functional theory (DFT). Here we explored the pyrolysis of CALF-20 and ZIF-8 using machine learned potentials and established potentials that can simulate high temperature decomposition at near DFT accuracy. Standard random sampling and two biased sampling techniques were tested in an effort to sample the phase space of average zinc coordination number and bond length. These biased sampling methods showed significant improvement over random sampling; the resulting models were able to successfully recreate the environments seen in a DFT simulation. Using this model, we then simulated a one nanosecond quench of CALF-20 and ZIF-8 at 1500 K and 1750 K, respectively. This gave atomistic details of how the MOF behaved at high temperatures including gas formation, changes in zinc coordination environment and decomposition of linkers. This demonstrates the potential of using MLPs to simulate complex, high temperature processes in MOFs to gain a better understanding of reactivity and predict the features needed for new catalytic materials.