Metal-Organic Polyhedra (MOPs) have exceptional potential for host-guest chemistry, but their discovery is hindered by the large combinatorial space of their building units. Computational screening offers a powerful workflow for efficiently screening large sets of MOPs, however, reliable results depend on accurate modelling of the geometries and properties of the MOPs. In this paper, we extend our previous workflow, which assembled computation-ready MOPs using purely geometric operations, by incorporating post-assembly computational modelling. We benchmark a range of methods, including machine learning interatomic potentials (MLIPs) and tight-binding DFT, against 85 experimentally resolved MOP structures. The results show that geometry optimisation significantly refines the initial assembled MOP structures in terms of cavity and pore properties when compared against experimental structures, with MLIPs found to achieve good accuracy and excellent convergence rates. We applied the most reliable method to the entire dataset, integrating the resulting data within The World Avatar through the OntoMOPs ontology and enabling natural language querying. Lastly, we demonstrate the utility of this refined dataset by screening for MOPs with potential to act as hosts for a urea guest molecule.
Butler et al. (Thu,) studied this question.