Toward a Generalizable Machine-Learned Potential for Metal–Organic Frameworks

Abstract

Machine-learned potentials (MLPs) have transformed the field of molecular simulations by scaling “quantum-accurate” potentials to linear time complexity. While they provide more accurate reproduction of physical properties as compared to empirical force fields, it is still computationally costly to generate their training data sets from ab initio calculations. Despite the emergence of foundational or general MLPs for organic molecules and dense materials, it is unexplored if one general MLP can be effectively developed for a wide variety of nanoporous metal–organic frameworks (MOFs) with different chemical moieties and geometric properties. Herein, by leveraging upon data-efficient equivariant MLPs, we demonstrate the possibility of developing a general MLP for nearly 3000 Zn-based MOFs. After curating a training data set comprising augmented MOF structures generated from density functional theory optimization, we validate the reliability of the general MLP in predicting accurate forces and energies when evaluated on a test set with chemically distinct MOF structures. Despite incurring slightly higher errors on structures containing rare chemical moieties, the general MLP can reliably reproduce physical (e.g., vibrational, thermodynamic, and mechanical) properties for a large sample of Zn-based MOFs. Crucially, by developing one MLP for many MOFs, the computational cost of high-throughput screening is potentially reduced by a few orders of magnitude. This enables us to predict quantum-accurate properties for notable Zn-MOFs that were previously uninvestigated via expensive theoretical calculations. To facilitate computational discovery among other families of complex chemical structures, we provide our data set and codes in the public Zenodo repository.