A new set of machine-learning models for metal-organic frameworks (MOFs) reported a sharp improvement in one of the study's central calculations: estimating adsorption enthalpy, a measure of how strongly a gas binds to a material. The uMOF models more than halved the error of their parent models, reaching a mean absolute error of 4 kJ/mol, and beat specialized gas-capture models by more than 80%. The paper describes that agreement as within experimental uncertainty.
The work is an arXiv v2 preprint dated 8 September 2026. It combines the models with a large DFT dataset, a literature-derived benchmark and tests of mechanical, vibrational and adsorption-related properties.
A training set built around many structures
The core uMOF corpus contains 85,524 DFT configurations spanning 19,950 unique MOFs and 79 elements. It was divided into 68,394 training frames, 8,470 validation frames and 8,660 test frames.
The final collection mixed 59,661 Gaussian-noise configurations with 12,621 geometry-optimization configurations, 5,987 molecular-dynamics configurations, 810 equation-of-state configurations and 5,646 adsorbate-loaded configurations. It also included 79 isolated atoms and 720 isolated gas molecules. A separate augmentation added 37,732 training frames and 4,192 validation frames, but the final models used the core h5_mof corpus alone after the authors observed lower downstream accuracy when the augmentation was included.
Reference energies, forces and stresses were generated with DFT at the r2 SCAN-D4 level, using VASP 6.4.2 through atomate2. The two models, uMOF-MH and uMOF-POLAR, were fine-tuned from MACE-MH-1 and MACE-POLAR-1-M. Both used isolated-atom references and a two-stage schedule; uMOF-POLAR also included explicit long-range electrostatics.
The clearest gains came in adsorption
On held-out validation data, uMOF-MH recorded an energy error of 33.2 meV per atom and a force error of 0.066 eV per angstrom. The corresponding figures for uMOF-POLAR were 39.4 meV per atom and 0.211 eV per angstrom.
An ablation test, with validation and test splits held fixed, removed 1,146 Cu-MOF molecular-dynamics frames, or 1.7% of the training set. The pre-SWA force RMSE stayed at 0.081 eV per angstrom, but the final force RMSE rose from 0.081 to 0.101 eV per angstrom; energy changed by +1.2 meV per atom.
The gains were not identical across every reported property. On the Tier-1 panel, bulk-modulus mean absolute error was 3.92 GPa for uMOF-MH versus 3.54 GPa for its parent, and 4.94 GPa for uMOF-POLAR versus 4.69 GPa for its parent. The reported phonon-derived heat-capacity errors were 8.59% and 7.59%, respectively.
A benchmark built from the literature
The literature benchmark contains 3,986 verified property values, including 3,146 experimental values, drawn from 626 papers and linked to more than 650 CIFs. Its versioned workflow processed MOF and COF papers, then checked extracted quotes deterministically in a window extending two chunks in either direction. A match had to reach a Levenshtein similarity of at least 0.95, and no further language-model call was used at verification.
In a seven-metal substitution series, uMOF-MH identified the experimentally expected open-metal binding site as the most stable for four of seven metals. Selected metal-CO2 distances and M-O-C angles were reproduced within average deviations of 0.06 angstrom and 4.5 degrees.
The isotherm test offered another reported gain, but in a tightly defined setting. In rigid GCMC simulations of CO2 adsorption for Mg-MOF-74 at 298 K, the uMOF models reproduced the experimental trajectory and high-pressure plateau, while the dispersion-corrected parent overshot and the uncorrected parent saturated prematurely. The comparison covered one framework and one gas under rigid GCMC, so it does not establish performance for flexible frameworks generally.
In a separate functional comparison, r2 SCAN-D4 produced 2% to 4% lower geometry errors than PBE-D3(BJ), relative to experiment, for the two tested frameworks. The difference was most pronounced for spin-polarized Cu-HKUST-1.
The evidence has a defined reach
The reported adsorption evidence comes from selected computational benchmarks: an adsorption-enthalpy comparison, a seven-metal binding-site series and one Mg-MOF-74 CO2 isotherm at 298 K. These tests do not establish universal performance across every MOF, gas or operating condition.
The dataset, benchmark, model weights, training and evaluation code, and benchmark pipeline were reported as publicly released through figshare and ml-peg. The work remains an arXiv v2 preprint dated 8 September 2026.
Paper data and sources
Original title: uMOF: A Universal Database, Benchmark, and Machine Learning Interatomic Potentials for Metal-Organic Frameworks
Authors: Théo Jaffrelot Inizan, Prathami Divakar Kamath, Alin Marin Elena, Kristin A. Persson
Journal/Repository: arXiv
Status: Preprint, not yet peer-reviewed
First online: 2026-08-28
DOI: Not available
Original paper · Full text