Machine learning interatomic potentials for amorphous zeolitic imidazolate frameworks

The detailed understanding of the microscopic structure of amorphous phases of metal-organic frameworks (MOFs) remains a widely open question: characterization of these systems is very difficult, both from the experimental and computational point of view. In molecular simulations, approaches have been proposed that rely either on reactive force field, that lack chemical accuracy, or first-principles calculations, that are too computationally expensive. Here, we have found an innovative solution to these problems by training a machine learning potential for the description of disordered phases of a zeolitic imidazolate framework (ZIF). We then used it to produce high-quality atomistic models of ZIF glasses, with accuracy close to density functional theory (DFT) but at far lower computational cost in production runs. Accurate microscopic models of amorphous metal-organic frameworks (MOFs) are difficult to create. Machine learning potentials based on data from ab initio molecular dynamics offer a novel way to achieve this goal.