uMOF: A Universal Database, Benchmark, and Machine-Learning Interatomic Potentials for Metal--Organic Frameworks

Foundation machine-learning interatomic potentials (MLIPs) deliver near-ab-initio accuracy at a fraction of the cost, but their promise for metal-organic frameworks (MOFs) remains largely unrealized as large unit cells make training data generation expensive with fine-tuned models and benchmarks all remaining scarce. We introduce uMOF: (i) the largest, most accurate DFT dataset for MOFs, at the r2SCAN-D4 level across 85,524 structures spanning 19,950 unique MOFs and 796 elements, for empty and gas-loaded MOFs; (ii) a literature-mined benchmark of 3,986 verified property (3,146 experimental), extracted from 626 papers by a multi-pass LLM pipeline with over 650 CIF structures; and (iii) two foundation MLIPs, the foundation MOF suite, uMOF-MH and uMOF-POLAR, fine-tuned from distinct MACE architectures on this dataset. While, a similar performance against state-of-the-art foundation models is observed on near-equilibrium properties (e.g., bulk modulus, phonon-derived heat capacity), on harder dynamical properties such as gas adsorption via Widom insertion and isotherm, uMOF models outperform the other models while closely matching experimental adsorption enthalpies and isotherm. We trace this to the quality and composition of the dataset and our benchmark, important for monitoring training. We found that including molecular dynamics trajectories is important for model stability and adsorbate-MOF interactions are better captured with r2SCAN-D4. We release the dataset, benchmark and models as a reproducible standard for MLIPs in reticular chemistry.