Toward\na Generalizable Machine-Learned Potential for\nMetal–Organic Frameworks

Machine-learned\npotentials (MLPs) have transformed the field of\nmolecular simulations by scaling “quantum-accurate”\npotentials to linear time complexity. While they provide more accurate\nreproduction of physical properties as compared to empirical force\nfields, it is still computationally costly to generate their training\ndata sets from ab initio calculations. Despite the emergence of foundational\nor general MLPs for organic molecules and dense materials, it is unexplored\nif one general MLP can be effectively developed for a wide variety\nof nanoporous metal–organic frameworks (MOFs) with different\nchemical moieties and geometric properties. Herein, by leveraging\nupon data-efficient equivariant MLPs, we demonstrate the possibility\nof developing a general MLP for nearly 3000 Zn-based MOFs. After curating\na training data set comprising augmented MOF structures generated\nfrom density functional theory optimization, we validate the reliability\nof the general MLP in predicting accurate forces and energies when\nevaluated on a test set with chemically distinct MOF structures. Despite\nincurring slightly higher errors on structures containing rare chemical\nmoieties, the general MLP can reliably reproduce physical (e.g., vibrational,\nthermodynamic, and mechanical) properties for a large sample of Zn-based\nMOFs. Crucially, by developing one MLP for many MOFs, the computational\ncost of high-throughput screening is potentially reduced by a few\norders of magnitude. This enables us to predict quantum-accurate properties\nfor notable Zn-MOFs that were previously uninvestigated via expensive\ntheoretical calculations. To facilitate computational discovery among\nother families of complex chemical structures, we provide our data\nset and codes in the public Zenodo repository.