Generalization of uMLIP accuracy to diverse MOFs and derived properties

Establish whether the agreement of universal machine learning interatomic potentials (uMLIPs) with ab initio energies and forces observed on selected metal–organic frameworks extends to a broader variety of metal–organic frameworks and to derived properties such as elastic moduli (bulk modulus), constant-volume heat capacity, and host–guest interaction energies.

Background

Prior studies report proof-of-principle results that uMLIPs can reproduce ab initio energies and forces for selected MOFs and yield stable molecular dynamics trajectories on specific structures. However, MOFs encompass vast chemical and structural diversity, and downstream properties often probe different regions of the potential energy surface than equilibrium fitting tasks.

The authors note that fine-tuning uMLIPs on MOF-specific data can improve performance, but narrows chemical coverage and compromises universality. This motivates a systematic benchmark across diverse MOFs and properties, explicitly testing whether favorable agreement generalizes beyond a small set of prototypical frameworks to complex derived properties relevant for practical applications.