betaFeedback

Reading v1View latest version

WorksPhysics and Astronomy

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

Abstract

Foundation machine learning interatomic potentials (MLIPs) deliver near-ab-initio accuracy at a fraction of the computational cost, yet their promise for Metal-organic Frameworks (MOFs) remains largely unrealized as large unit cells make first-principles training data expensive to generate, fine-tuned models are scarce, and experimentally grounded benchmarks are scarcer still. We introduce uMOF, a three-part contribution addressing this gap. First, we release the largest and most accurate density functional theory dataset for MOFs to date, computed at the r2^2SCAN-D4 level of theory across 85524 configurations spanning 19950 unique frameworks and 79 elements, covering empty and gas-loaded structures, geometry optimizations, equations of state, and finite-temperature molecular dynamics. Second, we release a literature-mined benchmark of 3986 verified property values (3146 experimental) extracted from 626 papers by a seven-stage, checkpointed multi-pass large language model pipeline, linked to more than 650 crystallographic information files. Third, we release two universal MLIPs for MOFs, uMOF-MH and uMOF-POLAR, fine-tuned from two architecturally distinct MACE foundation models on the uMOF dataset. On near-equilibrium, ``Tier-1'' properties (bulk modulus, phonon-derived heat capacity) the uMOF models perform comparably to existing foundation and fine-tuned baselines. On harder, dynamics-sensitive properties like gas adsorption enthalpies via Widom insertion and adsorption isotherms, the uMOF models outperform every baseline we test, including MOF-specialized gas-capture models trained on datasets up to three orders of magnitude larger, cutting error by more than 80% to within experimental uncertainty. We trace this advantage to the physical diversity of the training data and to level of theory where a small (1.7%) fraction of MD simulations is decisive for MLIP stability.

Introduction

Foundation machine-learning interatomic potentials (MLIPs) [1, 2, 3, 4, 5] have transformed atomistic simulation over the recent years, replacing expensive quantum-mechanical energy and force evaluations with learning surrogates that approach density functional theory (DFT) accuracy at orders of magnitude lower cost. Trained across a substantial fraction of the periodic table on thousands to millions of DFT data points [6, 7, 8, 9], these models promise transferability from a single pretrained checkpoint [10, 11, 12, 13], driving rapid adoption for crystal-structure prediction, high-throughput screening, and materials discovery. Equivariant message-passing backbones such as MACE [10] are among the most widely used, and while early foundation models were trained at the PBE or PBE+U level, more recent efforts have moved toward meta-GGA functionals [9, 14] such as r2^{2}SCAN [15], which reduce self-interaction and delocalization error relative to PBE [9, 12]. Despite this progress, training corpora remain dominated by dense inorganic bulk crystals and small molecules, leaving structurally distinct material classes underrepresented and their MLIP accuracy largely unevaluated.

Metal-organic frameworks (MOFs) [16, 17, 18] are a particularly demanding test case as a MOF unit cell couples inorganic metal nodes, organic linkers, and often guest molecules within porous, low-density lattices exceeding a thousand atoms per cell. The same chemistry that makes MOFs valuable for gas storage [19], separation [20], and catalysis [21] also makes them combinatorially diverse and physically distinct from the training distribution of most foundation models [?, 23]. Properties central to MOF applications like adsorption enthalpies, thermal expansion, framework flexibility are governed by long-range dispersion interactions [24, 25] and finite-temperature dynamics that short-range, near-equilibrium-trained foundation models are not designed to capture. Compounding this, because large unit cells make DFT single-point calculations expensive, MOF training data [?, 26, 27, 28] is built almost exclusively at the PBE [29] level, with dispersion corrections applied inconsistently or omitted, even though PBE is known to poorly capture spin-polarization in open-shell metal nodes [30] and the van der Waals interactions governing guest binding [31, 32].

Current MOF-specific MLIPs fall majorly into three families: active-learning, system-specific models [33, 34, 35, 36] that do not scale across chemical space; guest-adsorption-focused models [28, 37, 38, 39], which inherit PBE’s systematic mis-description of dispersion-dominated guest-framework interactions relative to higher-level functionals [32, 40]; and foundation-model fine-tunes like MACE-MP-MOF0 [?], a PBE+D3(BJ) [41, 42] fine-tune restricted to 24 elements and developed mainly for near-equilibrium properties, leaving guest-adsorption and high-temperature behavior insufficiently captured. On the benchmarking side, general-purpose suites like Matbench Discovery [13, 43] offer little MOF-specific coverage and primarily compare against DFT1 rather than curated experimental data at scale.

LLM-based literature mining [44] offers promise but so far is an underexplored route to scalable, experimentally grounded benchmarks.

This paper introduces uMOF, a dataset, benchmark, and model release designed to close these gaps simultaneously. We contribute: (i) the uMOF dataset (§2.1), the first r2^{2}SCAN-D4 [45] DFT dataset for MOFs at scale with 85,524 configurations from 19,950 frameworks spanning 79 elements, combining finite-temperature MD, relaxation/equation-of-state trajectories, and Gaussian noise perturbed structures, plus six gas adsorbates in industrially relevant empty and loaded MOFs, (ii) a literature-mined experimental benchmark (§2.2), 3,986 verified property records (3,146 experimental) across 20 property classes extracted from 626 papers via a multi-stage LLM workflow with deterministic hallucination filtering, linked to over 650 CIF structures and released as an open module of ml-peg [46], (iii) the uMOF models (§3), uMOF-MH and uMOF-POLAR, two architecturally distinct MACE [10] fine-tunes that substantially improve on their parent foundation models and on MACE-MP-MOF0 [?] on gas-adsorption properties; and (iv) a fine-tuning and data-generation protocol (§4.1) showing that finite-temperature MD, even at ~1.7% of the training set, is decisive for off-equilibrium force-model stability, and that level of theory matters more than raw data volume: uMOF outperforms gas-capture-specialized models like UMA-ODAC trained on ~1,000× more data by over 80% on adsorption enthalpy.

Methods

The uMOF Dataset

Level of theory

Every configuration in the uMOF dataset carries DFT energies, forces, and stresses computed at the r2^{2}SCAN-D4 level: the r2^{2}SCAN meta-GGA functional [47] improves systematically on PBE, while remaining tractable at the scale required for an 85,524-configuration dataset. The charge-dependent D4 dispersion correction [48] is applied throughout and, as we show quantitatively in Appendix A, benchmarking optimized geometries against experiment for a non-spin-polarized (Zn-MOF-5) and a spin-polarized (Cu-HKUST-1) framework shows r2^{2}SCAN-D4 gives 2–4% lower errors than PBE-D3(BJ) (Appendix Figure 4. These framework-level gains are consistent with broad molecular benchmarks, where r2^{2}SCAN-D4 ranks among the most accurate meta-GGAs across thermochemistry, transition-metal and metal-organic reactions [49], host-guest non-covalent interactions, and dispersion-bound molecular-crystal energies and equations of state [49, 50].

Benchmarking DFT functionals for MOF-5 Zn and HKUST-1 Cu

Figure 4. Benchmarking DFT functionals. r2r^2SCAN-D4 produces 2–4% lower errors than PBE-D3(BJ) relative to experiment for optimized geometries of Zn-MOF-5 and Cu-HKUST-1.

Overview of uMOF dataset, models, and benchmark

Figure 1. Overview of uMOF: a r2^{2}SCAN-D4 DFT dataset, two fine-tuned foundation MLIPs, and a literature-mined experimental benchmark that together bridge MLIP development and MOF experiments.

All DFT calculations were performed with VASP 6.4.2 [51, 52, 53, 54] using the atomate2 workflow manager [55]. Plane-wave cutoff energy (ENCUT) was selected by benchmarking 520, 680, 720, and 900 eV for cost versus accuracy in reproducing experimental unit-cell parameters and bulk moduli; we settled on ENCUT = 680 eV with EDIFF = 10−510^{-5} eV and LREAL=False for accurate forces. Spin-polarization was turned on (ISPIN=2) only for open-shell-metal-based MOFs as r2^{2}SCAN+D4 scales expensively with spin-polarization for large systems like MOFs. Remaining INCAR settings follow the MatPES r2^{2}SCAN static-calculation generator (MatPesMetaGGAStaticSetGenerator) in atomate2, consistent with the broader MatPES effort [9].

Sampling strategy and dataset composition

A dataset intended to train a transferable, dynamics-capable MOF potential must cover more than relaxed, near-equilibrium geometries. We combine three complementary sampling strategies (Table 1):

Data TypeUnique CountTotal Configurations
Gaussian noise19,922 MOFs59,661
Geometry optimization1,288 MOFs12,621
Molecular dynamics236 MOFs5,987
Strained, equation of state143 MOFs810
Adsorbate-loaded frameworks13 MOFs × 6 adsorbates5,646
Isolated atoms79 elements79
Isolated gas molecules6 adsorbates720
Total (distinct)19,95085,524

Table 1. Table 1: Composition of the uMOF dataset. The Molecular dynamics row aggregates three trajectory sets: Cu-MOF MD (1,433 frames, ∼70 Cu-based frameworks), temperature-sampled MOF MD (4,062 frames, ∼160 chemically varied frameworks spanning In, Zn, Mg, Al, and predominantly other closed-shell metal MOFs), and empty-framework MD (492 frames, 6 guest-free frameworks).

(i) Gaussian-noise perturbation seeded from QMOF structures [56, 57], which cheaply covers a broad range of chemistries at modest structural displacement (σ≈0.06\sigma\approx0.06 Å); (ii) geometry-relaxation and equation-of-state (EOS) trajectories, which cover the near-equilibrium potential-energy surface relevant to elastic and thermodynamic properties; and (iii) finite-temperature molecular dynamics, which is the only source of genuinely off-equilibrium, high-force configurations in the dataset spanning a wide range (mean ∣F∣|F| = 2.48 eV Å−1^{-1}, 95th percentile 7.60 eV Å−1^{-1}). These are supplemented with curated industrially relevant empty and adsorbate-loaded frameworks across six gas species, and isolated-atom and molecule references required for consistent energy referencing during training.

In total, the dataset spans 85,524 configurations across 19,950 unique MOFs and 79 elements (Z∈{1,3Z \in\{1, 3–9,119, 11–17,1917, 19–35,3735, 37–53,5553, 55–83,90,9283, 90, 92–94}94\}). The same framework can appear in more than one sampling category (e.g., both as a Gaussian-noise structure and as an MD trajectory seed), so the totals in Table 1 are distinct counts rather than column sums.

Training/validation/test splits.

The core training corpus from uMOF (h5_mof) contains 68,394 train, 8,470 validation and 8,660 test frames (6,823,332 / 847,345 / 859,897 atoms respectively), drawn from 21,519 distinct frameworks: 21,427 QMOF structures [58] plus roughly 90 common-name experimental MOFs (including UiO, MIL, ZIF, HKUST, MOF-74, IRMOF, CAU, and CALF-20 families). A separate Gaussian-noise augmentation set (h5_mofgen), generated from structures produced by our MOFGen diffusion model [59], adds 37,732 train and 4,192 validation frames (no held-out test split), consisting of DFT relaxation trajectories (≈12.5 frames per framework) for 3,350 generated frameworks disjoint from QMOF. Combined, the two sources total 127,448 configurations (12,885,556 atoms) across 24,869 unique frameworks. A neighbor cutoff of rmax⁡=4.5 A˚r_{\max} = 4.5\,\mathring{A} yields approximately 30 neighbors per atom on average across the dataset. As discussed in §3, we ultimately found that including the diffusion-generated augmentation set at training time degraded downstream accuracy, so the final released models are fine-tuned on the uMOF h5_mof alone (Table 1), while h5_mofgen is released alongside it for the community to use as they see fit.

The uMOF Benchmark

Assessing whether an MLIP is actually useful for MOF applications requires comparison against experiment, not just against DFT. But experimental MOF property measurements are scattered across thousands of papers, reported in inconsistent units, under varying conditions, and often only as figure data rather than tabulated values. Manually curating such a benchmark at scale is prohibitively labor-intensive. We instead build an automated multi-pass LLM pipeline to extract property values

directly from manually downloaded papers, with deterministic verification steps designed to suppress hallucination.

Pipeline design

The benchmark is produced by a multi-stage pipeline operating over MOF and COF papers, where each stage reads and writes a versioned JSON file so that any individual stage can be re-run or audited independently of the others.

  1. Load. Collect articles and candidate DOIs for which machine-readable text chunks are available.

  1. Rank (Claude Haiku 4.5). Score each paper’s title and abstract on a 0 to 5 relevance scale for simulation-relevant content, processed in batches of 20 papers.

  1. Extract (Claude Opus 4.7). For papers scoring above the threshold, emit one record per extracted property: a value, unit, measurement conditions (temperature, pressure, guest species), level of theory or experimental method, source location within the paper, a verbatim supporting quote with chunk index, a flag for whether the value is figure-only, and a reproducibility verdict covering feasibility, computational backend, workflow, and MLIP-compatibility.

  1. Verify. Deterministically fuzzy-match each extracted quote against its source chunk within a ±2-chunk window, requiring Levenshtein similarity ≥ 0.95; this stage makes no further LLM calls, so it cannot itself introduce new hallucinations.

  1. Score and filter. Aggregate a per-paper reproducibility score from the verified-quote fraction, presence of a code or CIF reference, and completeness of the reported method; drop properties flagged infeasible and drop entire papers with verified fraction below 0.6 or with no viable route to a CIF structure. The reproducibility score is a proxy for, not a guarantee of, correctness.

  1. Multimodal upgrade (Claude Opus 4.7). Reread figure-locked values directly from rendered figure images for up to 50 papers, recovering values that are visually but not textually reported.

  1. Export. Assemble the final benchmark.json file with manually added advanced entries like inelastic neutron scattering (INS) spectra compiled from literature review, together with a per-stage audit file recording record counts, drop reasons, API cost, model versions, configuration hash, and git commit SHA for full reproducibility.

Property tiers and dataset statistics

We organize the resulting 20 property classes into three tiers reflecting the cost of reproducing them with an MLIP calculator: Tier 1 covers static or finite-difference properties (lattice parameters, cohesive/formation energy, elastic tensor, bulk/shear/Young’s modulus, equation of state, heat capacity, INS spectra); Tier 2 covers short (≤1\le1 ns) molecular dynamics properties (NPT density, thermal expansion, self-diffusion, radial distribution function, NVT stability); and Tier 3 covers specialized-but-inexpensive properties centered on adsorption (Henry coefficients via Widom insertion, isosteric heat of adsorption at low loading, binding-site enumeration). The full breakdown is given in Table 2.

Property NameTierNo. of data pointsNo. of experimental data points
Lattice parameters129132656
QstQ_{st} (low loading)3283225
Binding energy122237
Binding-site enumeration310459
MSD / self-diffusion2715
Bulk modulus16920
NPT density26155
Cohesive energy1600
Henry coefficient35931
Thermal expansion23717
Formation energy1250
Young’s modulus11513
Elastic tensor1140
INS spectrum1127
Shear modulus1100
Cell vs. pressure188
Heat capacity177
Equation of state163
Radial distribution fn.261
NVT stability242
Total3,9863,146

Table 2. The uMOF benchmark by property class. Tier 1 = static/finite-difference, Tier 2 = short MD, Tier 3 = specialized (adsorption-centric).

The pipeline processed 626 candidate papers into 3,986 verified property records (79% experimental, 3,146 records) linked to more than 650 CIF structures drawn from CoRE-MOF 2025 [60] and MOSAEC-DB [61]. Structures are linked to their source papers by DOI 272 of 610 source papers (44.6%) resolve to a deposited CIF, yielding 830 candidate CIFs total.

The benchmark will also be released as an open module of ml-peg [46]1, independent of the specific models introduced in this paper, so that any MLIP can be scored against curated MOF experiments in a reproducible way.

Models and Fine-Tuning

We fine-tune two architecturally distinct MACE [10] foundation models on the uMOF dataset, chosen to probe whether an explicit long-range electrostatic term is beneficial for MOF chemistry beyond what a well-chosen dispersion correction and a higher level of theory already provide. (i) uMOF-MH fine-tunes the r2r^2SCAN-D4 MatPES head of MACE-MH-1, a multi-head foundation model with a short-range cutoff and no explicit long-range electrostatic term [12]. (ii) uMOF-POLAR fine-tunes MACE-POLAR-1-M, a polarizable variant of MACE that includes an explicit long-range electrostatic term conditioned on total charge and spin multiplicity [62], with elemental coverage up to Z≤83Z \le83.

Both models use isolated-atom energy references, a two-stage training schedule with stochastic weight averaging (SWA), and a force/stress-weighted loss in the second stage. uMOF-MH and uMOF-POLAR additionally retains a replay head from their foundation training to limit catastrophic forgetting of the broader chemical space seen during pretraining. For comparison against an existing PBE-level baseline, we also evaluate MACE-MP-MOF0 (v2) [?], a PBE+D3(BJ) MOF fine-tune of MACE-MP-0b, allowing us to isolate the effects of base architecture, training data, and level of theory on downstream accuracy. Table 3 summarizes the model panel and held-out training accuracy: both uMOF fine-tunes converge to small held-out force errors (≤0.2 eV A˚−1\le0.2\,\mathrm{eV}\,\mathrm{\mathring{A}}^{-1} RMSE), confirming that fine-tuning successfully transfers the foundation representation to r2r^2SCAN-D4 without affecting it.

ModelBase modelTraining data/headLong-rangeRMSE-EE (meV/atom)RMSE-FF (eV Å−1^{-1})
MACE-MH-1+D3foundationMatPES (r2^{2}SCAN)add-on——
uMOF-MHMACE-MH-1uMOF (r2^{2}SCAN-D4)from DFT33.20.066
MACE-POLAR-1-MfoundationOMol (ω\omegaB97M-V-rv10)yes——
uMOF-POLARMACE-POLAR-1-MuMOF (r2^{2}SCAN-D4)yes39.40.211
MACE-MP-MOF0-v2MACE-MP-0bMOF0 (PBE+D3(BJ))from DFT——

Table 3. Model panel and training accuracy. Held-out validation errors (best epoch, r2^{2}SCAN-D4 head, per-atom energy) are reported for the two uMOF fine-tunes

Results and Discussion

Molecular dynamics data is decisive for training stability

The molecular dynamics portion of the uMOF dataset (5,987 frames total, Table 1) aggregates three trajectory sets: Cu-MOF MD set (1,433 frames across roughly 70 Cu-based frameworks as Cu is an abundant and open-shell metal that is challenging to model in MOFs), other temperature-sampled MOF

MD (4,062 frames across roughly 160 chemically varied, predominantly closed-shell-metal frameworks spanning Zr, Zn, Mg, and Al), and empty-framework MD (492 frames across 6 guest-free frameworks). To isolate the contribution of finite-temperature sampling, we removed only the Cu-MOF MD subset (1,146 training frames, 1.7% of the training set) from an otherwise identical MACE-MH-1 fine-tuning recipe, holding validation and test splits fixed.

Even this small removal has an outsized effect (Table 4). The two runs are essentially identical through the first training stage but once the force/stress-weighted SWA stage begins, the reduced-data model’s validation force error spikes to 0.154 eV Å−1^{-1} and settles roughly 25% worse than the

Training dataFramesRMSE-EE [meV/atom]RMSE-FF pre-SWA [eV Å−1^{-1}]RMSE-FF final [eV Å−1^{-1}]
With Cu-MOF MD68,39437.30.0810.081
Without Cu-MOF MD67,24838.50.0810.101

Table 4. MD ablation (identical MACE-MH-1 fine-tune recipe). Removing the Cu-MOF MD subset (1,146 frames, 1.7% of the training set) leaves pre-SWA force accuracy unchanged and energy essentially unchanged (+1.2 meV/atom), but degrades the force-weighted SWA stage by ∼25%.

full-data model (0.101 vs. 0.081 eV Å−1^{-1} final RMSE-FF), while energy accuracy is essentially unchanged (+1.2 meV/atom).

We re-evaluated both stage-two models on a per-configuration-type basis across the full validation set (Table 5). The degradation is not confined to the held-out MD frames: the QMOF Gaussian-noise single-point structures, which make up 70% of the validation set and are chemically distinct from the removed Cu-based frameworks, degrade by the same ∼25%\sim25\%. Near-equilibrium categories, geometry optimization, equation of state, MOF0-derived structures, adsorbate-loaded frameworks, and empty frameworks barely move (+3+3–4%4\%). Both finite-temperature MD snapshots and Gaussian-noise-perturbed statics populate the off-equilibrium, high-force regime of the potential-energy surface and the Cu-MOF MD data teaches the model exactly this regime. The force-weighted SWA stage relies on that signal precisely where the bulk of the QMOF Gaussian-noise data lives. Molecular dynamics is the only part of the corpus that samples the high-force regime governing dynamics, Widom insertion, and anharmonic phonon calculations downstream, so even a 1.7% Cu-MOF MD slice functions as a decisive, non-optional ingredient of MOF fine-tuning data rather than a minor augmentation.

Configuration typenn (val)With Cu-MOF MDWithoutΔ\Delta
QMOF Gaussian noise (single-point)5,9670.0900.113+25%
MD (Cu-MOF)1430.0740.112+51%
Equation of state1160.0650.068+4%
Geometry optimization1,2530.0540.055+3%
MOF0-derived3010.0400.041+4%
Adsorbate-loaded5470.0250.026+4%
Empty frameworks710.0200.021+4%
Isolated molecules720.0120.021+73%
All8,4700.0810.101+25%

Table 5. Per-configuration-type validation force RMSE (eV Å−1^{-1}) for the MD ablation of Table 4. Removing the Cu-MOF MD subset degrades not only the held-out MD frames but the chemically distinct QMOF Gaussian-noise structures (70% of validation) by the same +25%, while near-equilibrium categories barely change – evidence that finite-temperature sampling transfers to the broader off-equilibrium force landscape.

Preserving accuracy on near-equilibrium mechanical and thermal properties

We first evaluate near-equilibrium, Tier-1 properties across nine benchmark MOFs spanning seven metal nodes and diverse linker chemistries (structures and experimental references listed in Appendix B). MACE-MH-1 without any dispersion correction achieves the lowest bulk-modulus error of any model tested (Table 6), consistent with the tendency of D3/D4-type corrections to overestimate long-range interactions in MOFs [63]. MACE-MP-MOF0, which is fine-tuned on a smaller, more targeted PBE+D3(BJ) dataset that overlaps heavily with the benchmarked structures, performs comparably. The uMOF models, fine-tuned on a dataset two orders of magnitude larger and spanning 79 vs. 24 elements, trade at most 0.04 GPa of bulk-modulus accuracy for substantially broader elemental and structural coverage. On this metric taken alone, model choice looks largely unaffected.

ModelBulk modulus MAE [GPa]MPAECv_{C_v} (%)% Imaginary modesWasserstein dist. [meV]
MACE-MH-13.547.220.393.25
uMOF-MH3.928.590.253.18
MACE-POLAR-1-M4.697.880.382.86
uMOF-POLAR4.947.590.283.27
MACE-MP-MOF0-v23.889.720.002.98

Table 6. Tier-1 property accuracy. Bulk-modulus MAE (nine MOFs); phonon-derived CvC_v error at 300 K, % imaginary modes, and INS Wasserstein distance at 10 K (seven MOFs).

Phonon-derived heat capacity (CvC_v), computed alongside the fraction of spurious imaginary phonon modes, tells a more cautionary story. MACE-MH-1 without dispersion again appears the most accurate model on aggregate CvC_v error, but this is an artifact as it produces the largest fraction of spurious imaginary modes, which are discarded before computing CvC_v and thereby silently remove exactly the contributions that would otherwise drive the metric toward larger, more realistic errors. This rewards an unstable potential and penalizes MACE-MP-MOF0, the only model in our panel with 0.00% imaginary modes with the largest apparent CvC_v error, precisely because it alone produces a physically complete vibrational spectrum within the defined convergence criteria (Appendix C). Comparison INS spectra, an independent experimental observable, corroborates this reading where MACE-MH-1’s low CvC_v error coexists with one of the largest INS Wasserstein distances from experiment due to the missed spurious imaginary modes (Appendix Figure 5), confirming that its apparent CvC_v accuracy is an artifact of discarded modes rather than a genuinely correct vibrational spectrum. This echoes and independently corroborates the findings in [64] that aggregate accuracy metrics can reward models for the right answer for the wrong reason unless imaginary-mode fraction is explicitly audited; we recommend this auditing as standard practice for MLIP vibrational-property benchmarking going forward.

Overall, Tier-1 analysis indicates that the uMOF models achieve accuracy comparable to existing foundation and fine-tuned baselines on generic, near-equilibrium properties.

Gas adsorption: where uMOF models clearly pull ahead

Tier-1 tasks understate what matters for real MOF applications. Gas adsorption is one of the most widely studied MOF applications, and accurately predicting adsorption enthalpies and binding sites remains a central, long-standing open challenge for MLIPs [67]. Unlike bulk modulus and harmonic phonons, which probe the near-equilibrium potential-energy surface, adsorption requires jointly and correctly capturing both framework dynamics and guest-host dispersion interactions which is exactly the regime that §4.1 shows is governed by the MD fraction of the training data.

Using the flames package [24], we run Widom-insertion calculations at 298 K for 20,000 steps (convergence checked against step count on HKUST-1 as a representative case; full parameters in Appendix C) across 13 MOFs for CO2_2 and compare predicted binding enthalpies against experiment. We further test adsorption of N2_2, CH4_4, and H2_2 to test adsorbate sensitivity and accuracy in HKUST-1, a challenging test case. MACE-MP-MOF0 is excluded from this comparison, as it was fine-tuned specifically for phonon accuracy and is not suited to gas-adsorption calculations. Instead, other domain-specific models like UMA-ODAC [1, 27] and MACE-DAC [28] trained specifically for MOF gas-adsorption applications are used for comparison.

Binding-site characteristics.

The uMOF models are clearly and consistently superior to their corresponding foundation-model baseline (Figure 2 a), more than halving error to a mean absolute error (MAE) of 4 kJ/mol effectively reproducing experimental values considering experimental uncertainty for these systems [68]. Dispersion treatment matters throughout for MOFs where MACE-MH-1 without any dispersion correction significantly underestimates adsorption enthalpy, while adding D3 leads to significant overestimation, so neither variant achieves a good balance. MACE-POLAR-1, despite including an explicit long-range electrostatic term, performs worst of all baselines tested, a counterintuitive result. Fine-tuning on uMOF’s r2^2SCAN-D4 data resolves this balance for both architectures, indicating that the gain comes from the level of theory and the composition of the training data rather than from architectural choices. The uMOF models surpass domain-specialized models trained explicitly for gas-capture applications by more than 80% on adsorption enthalpy MAE. UMA-ODAC25 is trained on ODAC2025, a PBE+D3 dataset of roughly 70 million MOF-adsorbate configurations which is nearly three orders of magnitude larger than uMOF’s training corpus. MACE-DAC is trained on the smaller, related GoldDAC dataset. That a dataset almost 1,000× smaller than ODAC2025 outperforms models purpose-built for gas capture at that scale is a direct demonstration that data volume alone does not guarantee downstream accuracy for this class of properties and level of theory matters at least as much, if not more. uMOF’s r2^2SCAN-D4 labels move beyond the GGA-level PBE used throughout existing MOF datasets, and this accuracy advantage persists despite uMOF’s relatively far smaller scale.

Model accuracy on experimental binding enthalpies from Widom insertion

Figure 2. Model accuracy on experimental binding enthalpies from Widom insertion: (a) uMOF-MH and uMOF-POLAR models outperform their pre-trained foundation models by more than 2 times. (b) The uMOF models show a clear advantage, with errors improved by more than 80% over other gas-capture-focused models for MOFs such as UMA-ODAC25 and MACE-DAC. Values for UMA-ODAC25 and MACE-DAC were extracted from Oliveira et al. [24], plotted with a standard deviation of 1 kJ/mol. Experimental values are taken from Queen et al. [65] and Grajciar et al. [66].

Binding-site characteristics.

Accuracy on adsorption enthalpy is not achieved by chance as we further evaluate binding-site characteristics across the M-MOF-74 series (M = Mg, Fe, Co, Ni, Mn, Cu, Zn), where CO2_2 is experimentally known to bind at the open metal site [69]. uMOF-MH identifies the correct open-metal binding site as the most stable for 4 of 7 metals (Mg, Ni, Mn, Zn); for M = Fe, Co, and Cu, the model instead places the molecule in the pore. For the correctly identified sites, the most accurate uMOF-MH reproduces the experimental M–O(CO2_2) bond distance and M–O–C bond angle to within 0.06 Å and 4.5° on average (Table 7, and the model reproduces the experimental trend of adsorption strength (Mg > Ni > Co > Fe > Mn > Cu,) across this metal-substitution series . Zn-MOF-74 is the only variant for which the uMOF-Mh model fails to get the right trend.

Metal (M)∣Δdistance∣\lvert\Delta_{\mathrm{distance}}\rvert [Å]∣Δangle∣\lvert\Delta_{\mathrm{angle}}\rvert [°]
Mg0.114.2
Ni0.033.5
Mn0.0053.2
Zn0.1017.18
MAE0.064.52

Table 7. uMOF-MH errors relative to experiment for M–O(CO2_2) bond distance and M–O–C bond angle across correctly identified binding sites in the M-MOF-74 series (M = Mg, Ni, Mn, Zn). Experimental references from Queen et al. [65].

Adsorption isotherms.

Beyond enthalpies, the shape of adsorption isotherm is a key observable, encoding the adsorption mechanism, pore-filling behavior and thermodynamic response across pressure that are critical for reticular chemistry. We performed rigid Grand Canonical Monte Carlo (GCMC) simulation on Mg-MOF-74 for CO2_2 isotherms with each model. We found that adding a D3 correction to MACE-MH-1 overshoots the uptake (plateauing near 17.5 mmol g−1^{-1}) through

an overestimated adsorption enthalpy, whereas removing dispersion moves closer to experiment only through error compensation while still getting an incorrect shape. Our uMOF models, trained directly on r2^{2}SCAN-D4 data with dispersion built in, instead reproduce the experimental isotherm shape with a slight plateau shift while the parent foundations saturate prematurely or are significantly off Figure 3. The difference is a small systematic shift that additional CO2_2-MOF training data should close. This agreement over the full isotherm shows that uMOF captures the broader adsorption physics required for reliable MOF screening.

Simulated CO2 adsorption isotherms for Mg-MOF-74

Figure 3. Simulated CO2CO_{2} adsorption isotherms for Mg-MOF-74 at 298 K (dual-site Langmuir fits to GCMC uptake) versus experiment. The uMOF models reproduce the experimental trajectory including the high-pressure plateau, whereas the dispersion-corrected foundation model overshoots and the uncorrected one saturates prematurely.

Conclusion

uMOF pairs the largest and most accurate DFT dataset for MOF chemistry to date (85,524 r2^{2}SCAND4 configurations, 19,950 frameworks, 79 elements) with an LLM-mined experimental benchmark (3,986 records, 20 property classes, distributed through the open ml-peg platform) and two fine-tuned MACE models, all released for community use. On generic near-equilibrium "Tier-1" properties the uMOF models match existing foundation and fine-tuned baselines. On gas adsorption, a substantially harder, dynamical task and a long-standing issue for MOF MLIPs, they show a clear, consistent advantage over every baseline tested, with predicted binding sites, geometries and isotherm that agree with experiment. This advantage traces to two complementary factors: a small but decisive fraction of finite-temperature MD trajectories that stabilizes MLIP out-of-equilibrium, and the r2^{2}SCAN-D4 level of theory, which better captures MOF-guest dynamics than the classic PBE-based data underlying prior large-scale MOF datasets. uMOF’s element-general representation should extend naturally to related reticular materials such as covalent-organic frameworks and polymers. Natural next step is broadening the benchmark’s Tier-2 dynamical coverage and material classes beyond MOFs.

Data and Code Availability

The uMOF dataset, benchmark, and model weights, together with training and evaluation code, are released publicly on figshare here: 10.6084/m9.figshare.33311829). The benchmark pipeline is distributed as a module of ml-peg at https://github.com/ddmms/ml-peg.

Acknowledgments

This research used computing resources of Lawrencium, National Energy Research Scientific Computing Center (NERSC), a Department of Energy (DOE) Office of Science User Facility and by Science and Technology Facilities Council Scientific Computing Department’s SCARF cluster. T.J.I acknowledges financial support from the Bakar Institute of Digital Materials for the Planet. P.D.K. acknowledges financial support from U.S. National Science Foundation’s "The Quantum Sensing Challenges for Transformational Advances in Quantum Systems" program No. 2326838. A.M.E. was also supported by the Ada Lovelace Centre at the Science and Technology Facilities Council (https://adalovelacecentre.ac.uk/), the Physical Sciences Data Infrastructure (https://psdi.ac.uk; jointly STFC and the University of Southampton) under grants EP/X032663/1 and EP/X032701/1, and EPSRC under grants EP/W026775/1 and EP/V028537/1.

DFT functional benchmarking against experiment

Since current large-scale MOF MLIP training data is overwhelmingly built at the PBE+D3(BJ) level, we benchmarked DFT-optimized geometries with r2r^2SCAN-D4 and PBE-D3(BJ) relative to experiment for two representative MOFs: Zn-MOF-5 (non-spin-polarized) and Cu-HKUST-1 (spin-polarized). r2r^2SCAN-D4 produces 2–4% lower geometry errors than PBE-D3(BJ) across both frameworks (Figure 4), with the improvement most pronounced for spin-polarized Cu-HKUST-1, where uncorrected PBE’s larger delocalization error leads to overestimated cell volumes relative to experiment [15]. The trade-off is a modest overestimation of bulk modulus and a blue-shift in phonon frequencies [70], consistent with the slightly higher bulk-modulus errors and blue-shifted INS spectra observed for r2r^2SCAN-D4-based MLIPs in Table 6 and Figure 5.

Predicted vs. experimental inelastic neutron scattering spectra for six representative frameworks

Figure 5. Predicted vs. experimental inelastic neutron scattering (INS) spectra for six representative frameworks: (a) MOF-5, (b) ZIF-8, (c) MIL-53, (d) MIL-140A, (e) ZIF-4, (f) ZIF-7. Experimental data from Lock et al. [80], Liu et al. [81], Ryder et al. [82], and Ryder et al. [83].

Benchmark structure selection

Bulk modulus. Nine MOFs spanning seven unique metal nodes, with experimental references: Cu-HKUST-1 [71], Ti-MIL-125 [72], Al-MIL-53 [73], Zr-UiO-66, Hf-UiO-66, and Ce-UiO-66 [74], Zr-UiO-67, Zn-ZIF-8 [75], and Zn-ZIF-4 [76].

Heat capacity. Seven MOFs, with experimental references: Zn-MOF-5, Cu-MOF-177, Cu-HKUST-1, and Al-MIL-53 [77], Zr-UiO-66 [78], and Zn-MOF-74 and Zn-ZIF-8 [79].

Inelastic neutron scattering (INS). Six MOFs, with experimental references: Zn-MOF-5 [80], Al-MIL-53 [81], Zn-ZIF-8, Zn-ZIF-4, and Zn-ZIF-7 [82], and Zr-MIL-140A [83]. The ZIF series primarily differ in the functional groups and side chains attached to the base imidazole linker, providing a sensitive probe of model accuracy for the effect of linker functionalization on framework dynamics (Figure 5).

Calculation parameters for benchmark evaluation

Widom Insertion. 13 MOFs, with experimental references: Zn-CALF-20 [84], Cu-HKUST-1 [85], Zn-MOF-5 [86], Al-MIL-53 [87], Zn-ZIF-8 [88], Zr-UiO-66 [89], and the seven M-MOF-74 series [65] for CO2_2 adsorption. With three additional points where HKUST-1 was used for methane, nitrogen and hydrogen adsorption calculations. [85].

Bulk modulus. Full-cell LBFGS optimization to fmax⁡=10−6 eV/A˚f_{\max} = 10^{-6}\,\mathrm{eV}/\mathring{\mathrm{A}}, followed by fixed-cell optimization at the same convergence for six strained configurations (±3%\pm3\% linear volumetric strain) using the janus package [90]; bulk modulus is obtained from a seven-point Birch–Murnaghan equation-of-state fit about the relaxed volume.

Heat capacity. Harmonic phonons computed from 0 to 1000 K on an 11×11×1111 \times11 \times11 mesh with Phonopy [91, 92], following full-cell LBFGS optimization to fmax = 10−8 eV/A˚10^{-8}\,\mathrm{eV}/\mathring{\mathrm{A}} or 1000 steps, whichever is reached first.

INS spectra. Computed using the workflow of Kamath et al. [93] with matching settings.

Enthalpy of adsorption. Widom-insertion calculations performed with the flames package [24] at 298 K for 20,000 steps; convergence with respect to step count was checked on HKUST-1 as a representative example, with 20,000 steps selected as balancing cost and accuracy.

References

  1. [1]Brandon Wood, Misko Dzamba, Xiang Fu, Meng Gao, Muhammed Shuaibi, Luis Barroso-Luque, Kareem Abdelmaqsoud, Vahe Gharakhanyan, John Kitchin, Daniel Levine, et al. Uma: A family of universal models for atoms. Advances in Neural Information Processing Systems, 38:129391–129427, 2026.
  2. [2]Mark Neumann, James Gin, Benjamin Rhodes, Steven Bennett, Zhiyi Li, Hitarth Choubisa, Arthur Hussey, and Jonathan Godwin. Orb: A fast, scalable neural network potential. arXiv preprint arXiv:2410.22570, 2024.arxiv.org/abs/2410.22570
  3. [3]Bowen Deng, Peichen Zhong, KyuJung Jun, Janosh Riebesell, Kevin Han, Christopher J Bartel, and Gerbrand Ceder. Chgnet as a pre-trained universal neural network potential for charge-informed atomistic modelling. Nature Machine Intelligence, 5(9):1031–1041, 2023.
  4. [4]Ilyes Batatia, Philipp Benner, Yuan Chiang, Alin M Elena, Dávid P Kovács, Janosh Riebesell, Xavier R Advincula, Mark Asta, Matthew Avaylon, William J Baldwin, et al. A foundation model for atomistic materials chemistry. The Journal of chemical physics, 163(18), 2025.
  5. [5]Chi Chen and Shyue Ping Ong. A universal graph deep learning interatomic potential for the periodic table. Nature Computational Science, 2(11):718–728, 2022.DOI
  6. [6]Deng B., Zhong P., Jun K., Riebesell J., Han K., Bartel C. J., and Ceder G. Chgnet as a pretrained universal neural network potential for charge-informed atomistic modelling. Nat. Mach. Intell., 5:1031–1041, 2023. doi: 10.1038/s42256-023-00716-3.DOI
  7. [7]Luis Barros-Luque, Muhammed Shuaibi, Xiang Fu, Brandon M Wood, Misko Dzamba, Meng Gao, Ammar Rizvi, Matt Uyttendaele, C Lawrence Zitnick, and Zachary W Ulissi. The open materials 2024 (omat24) inorganic materials dataset and models. Nature Computational Science, pages 1–11, 2026.
  8. [8]Mohammad M Ghahremanpour, Paul J Van Maaren, and David Van Der Spoel. The alexandria library, a quantum-chemical database of molecular properties for force field development. Scientific data, 5(1):180062, 2018.
  9. [9]Aaron D. Kaplan et al. MatPES: A foundational potential energy surface dataset for materials. arXiv preprint arXiv:2503.04070, 2025.arxiv.org/abs/2503.04070
  10. [10]Ilyes Batatia, Dávid P. Kovács, Gregor N. C. Simm, Christoph Ortner, and Gábor Csányi. MACE: Higher order equivariant message passing neural networks for fast and accurate force fields. In Advances in Neural Information Processing Systems (NeurIPS), 2022.arxiv.org/abs/2206.07697
  11. [11]Amil Merchant, Simon Batzner, Samuel S. Schoenholz, Muratahan Aykol, Gowoon Cheon, and Ekin Dogus Cubuk. Scaling deep learning for materials discovery. Nature, 624:80–85, 2023.
  12. [12]Ilyes Batatia, Philipp Benner, Yuan Chiang, et al. A foundation model for atomistic materials chemistry. arXiv preprint arXiv:2401.00096, 2024.arxiv.org/abs/2401.00096
  13. [13]Janosh Riebesell, Rhys E. A. Goodall, Philipp Benner, et al. Matbench discovery: A framework to evaluate machine learning crystal stability predictions. arXiv preprint arXiv:2308.14920, 2024.arxiv.org/abs/2308.14920
  14. [14]Matthew C Kuner, Aaron D Kaplan, Kristin A Persson, Mark Asta, and Daryl C Chrzan. Mp-aloe: an r2scan dataset for universal machine learning interatomic potentials. npj Computational Materials, 11(1):352, 2025.arxiv.org/abs/2507.05559
  15. [15]Furness J. W., Kaplan A. D., Ning J., Perdew J. P., and Sun J. Accurate and numerically efficient r2scan meta-generalized gradient approximation. J. Phys. Chem. Lett., 11:8208–8215, 2020. doi: 10.1021/acs.jpclett.0c02405.DOI
  16. [16]H.-C. Zhou, J. R. Long, and O. M. Yaghi. Introduction to metal–organic frameworks. Chem. Rev., 112:673–674, 2012. doi: 10.1021/cr200018s.
  17. [17]Hiroyasu Furukawa, Kyle E. Cordova, Michael O’Keeffe, and Omar M. Yaghi. The chemistry and applications of metal-organic frameworks. Science, 341(6149):1230444, 2013. doi: 10.1126/science.1230444. URL https://www.science.org/doi/abs/10.1126/science.1230444.DOI
  18. [18]Omar M. Yaghi, Michael O’Keeffe, Nathan W. Ockwig, Hee K. Chae, Mohamed Eddaoudi, and Jaheon Kim. Reticular synthesis and the design of new materials. Nature, 423(6941):705–714, 2003. doi: 10.1038/nature01650. URL https://doi.org/10.1038/nature01650.
  19. [19]Hao Li, Kecheng Wang, Yujia Sun, Christina T Lollar, Jialuo Li, and Hong-Cai Zhou. Recent advances in gas storage and separation using metal–organic frameworks. Materials Today, 21(2):108–121, 2018.
  20. [20]Jian-Rong Li, Julian Sculley, and Hong-Cai Zhou. Metal–organic frameworks for separations. Chemical reviews, 112(2):869–932, 2012.
  21. [21]JeongYong Lee, Omar K Farha, John Roberts, Karl A Scheidt, SonBinh T Nguyen, and Joseph T Hupp. Metal–organic framework materials as catalysts. Chemical Society Reviews, 38(5):1450–1459, 2009. [22] A. M. Elena, P. D. Kamath, T. Jaffrelot Inzian, A. S. Rosen, F. Zanca, and K. A. Persson. Machine learned potential for high-throughput phonon calculations of metal–organic frameworks. npj Comput. Mater., 11:125, 2025. doi: 10.1038/s41524-025-01611-8.
  22. [23]Hendrik Kraß, Ju Huang, and Seyed Mohamad Moosavi. Mofs imbench: evaluating universal machine learning interatomic potentials in metal–organic framework molecular modeling. npj Comput. Mater., 12(1):4, 2025. doi: 10.1038/s41524-025-01872-3. URL https://doi.org/10.1038/s41524-025-01872-3.
  23. [24]Felipe L Oliveira, Holger-Dietrich Saßnick, and Guillaume Maurin. Flames–a flexible and extensible code for monte carlo simulations of nanoporous materials. 2026.DOI
  24. [25]J. Edzards, J. Santana-Andreo, H.-D. Saßnick, and C. Cocchi. Benchmarking selected density functionals and dispersion corrections for mof-5 and its derivatives. J. Chem. Theory Comput., 21:7062–7074, 2025. doi: 10.1021/acs.jctc.5c00399.DOI
  25. [26]Sriram A., Choi S., Yu X., Brabson L. M., Das A., Ulissi Z., Uyttendaele M., Medford A. J., and Sholl D. S. The open dac 2023 dataset and challenges for sorbent discovery in direct air capture. ACS Cent. Sci., 10:923–941, 2024. doi: 10.1021/acscentsci.3c01629.DOI
  26. [27]Anuroop Sriram, Logan M Brabson, Xiaohan Yu, Sihoon Choi, Kareem Abdelmaqsoud, Elias Moubarak, Pim de Haan, Sindy Löwe, Johann Brehmer, John R Kitchin, et al. The open dac 2025 dataset for sorbent discovery in direct air capture. arXiv preprint arXiv:2508.03162, 2025.arxiv.org/abs/2508.03162
  27. [28]Yunsung Lim, Hyunsoo Park, Aron Walsh, and Jihan Kim. Accelerating co2 direct air capture screening for metal-organic frameworks with a transferable machine learning force field. Matter, 8(7), 2025.
  28. [29]John P. Perdew, Kieron Burke, and Matthias Ernzerhof. Generalized gradient approximation made simple. Phys. Rev. Lett., 77:3865–3868, Oct 1996. doi: 10.1103/PhysRevLett.77.3865. URL https://link.aps.org/doi/10.1103/PhysRevLett.77.3865.DOI
  29. [30]Rosen A. S., Notestein J. M., and Snurr R. Q. Comparing gga, gga+u, and meta-gga functionals for redox-dependent binding at open metal sites in metal–organic frameworks. J. Chem. Phys., 152:224101, 2020. doi: 10.1063/5.0010166.DOI
  30. [31]Florian R Rehak, GiovanniMaria Piccini, Maristella Alessio, and Joachim Sauer. Including dispersion in density functional theory for adsorption on flat oxide surfaces, in metal–organic frameworks and in acidic zeolites. Physical Chemistry Chemical Physics, 22(14):7577–7585, 2020.
  31. [32]Bess Vlaisavljevich, Johanna Huck, Zeric Hulvey, Kyuho Lee, Jarad A Mason, Jeffrey B Neaton, Jeffrey R Long, Craig M Brown, Dario Alfè, Angelos Michaelides, et al. Performance of van der waals corrected functionals for guest adsorption in the m2 (dobdc) metal–organic frameworks. The journal of physical chemistry. A, 121(21):4139, 2017.
  32. [33]Marco Eckhoff and Jörg Behler. From molecular fragments to the bulk: Development of a neural network potential for MOF-5. Journal of Chemical Theory and Computation, 15(6):3793–3809, 2019. doi: 10.1021/acs.jctc.8b01288.DOI
  33. [34]Sander Vandenhaute, Maarten Cools-Ceuppens, Simon DeKeyser, Toon Verstraelen, and Veronique Van Speybroeck. Machine learning potentials for metal–organic frameworks using an incremental learning approach. npj Computational Materials, 9(1):19, 2023. doi: 10.1038/s41524-023-00969-x.DOI
  34. [35]Sandro Wieser and Egbert Zojer. Machine learned force-fields for an ab-initio quality description of metal–organic frameworks. npj Computational Materials, 10(1):18, 2024. doi: 10.1038/s41524-024-01205-w.DOI
  35. [36]Ankit Sharma and Stefano Sanvito. Quantum-accurate machine learning potentials for metal–organic frameworks using temperature driven active learning. npj Computational Materials, 10(1):237, 2024. doi: 10.1038/s41524-024-01427-y.DOI
  36. [37]Ruben Goeminne, Louis Vanduyfhuys, Veronique Van Speybroeck, and Toon Verstraelen. DFT-quality adoption simulations in metal–organic frameworks enabled by machine learning potentials. Journal of Chemical Theory and Computation, 19(18):6313–6325, 2023. doi: 10.1021/acs.jctc.3c00495.DOI
  37. [38]Shanping Liu, Romain Dupuis, Dong Fan, Salma Benzaria, Mickaele Bonneau, Prashant Bhatt, Mohamed Eddaoudi, and Guillaume Maurin. Machine learning potential for modelling H₂ adsorption/diffusion in MOFs with open metal sites. Chemical Science, 15(14):5294–5302, 2024. doi: 10.1039/d3sc05612k.DOI
  38. [39]Anuroop Sriram, Sihoon Choi, Xiaohan Yu, Logan M. Brabson, Abhishek Das, Zachary Ulissi, Matt Uyttendaele, Andrew J. Medford, and David S. Sholl. The open DAC 2023 dataset and challenges for sorbent discovery in direct air capture. ACS Central Science, 10(5):923–941, 2024. doi: 10.1021/acscentsci.3c01629.DOI
  39. [40]Yeongsu Cho and Heather J Kulik. Improving gas adsorption modeling for mofs by local calibration of hubbard u parameters. The Journal of chemical physics, 160(15), 2024.
  40. [41]Stefan Grimme, Jens Antony, Stephan Ehrlich, and Helge Krieg. A consistent and accurate ab initio parametrization of density functional dispersion correction (DFT-D) for the 94 elements H-Pu. J. Chem. Phys., 132(15):154104, 04 2010. ISSN 0021-9606. doi: 10.1063/1.3382344. URL https://doi.org/10.1063/1.3382344.
  41. [42]Stefan Grimme, Stephan Ehrlich, and Lars Goerigk. Effect of the damping function in dispersion corrected density functional theory. J. Comput. Chem., 32(7):1456–1465, 2011. doi: https://doi.org/10.1002/jcc.21759. URL https://onlinelibrary.wiley.com/doi/abs/10.1002/jcc.21759.
  42. [43]Yuan Chiang, Tobias Kreiman, Christine Zhang, Matthew Kuner, Elizabeth Weaver, Ishan Amin, Hyunsoo Park, Yunsung Lim, Jihan Kim, Daryl Chrzan, et al. Ml ip arena: advancing fairness and transparency in machine learning inter-atomic potentials via an open, accessible benchmark platform. Advances in Neural Information Processing Systems, 38, 2026.
  43. [44]Yeonghun Kang, Wonseok Lee, Taeun Bae, Seunghee Han, Huiwon Jang, and Jihan Kim. Harnessing large language models to collect and analyze metal–organic framework property data set. Journal of the American Chemical Society, 147(5):3943–3958, 2025.
  44. [45]Eike Caldeweyher, Sebastian Ehlert, Andreas Hansen, Hagen Neugebauer, Sebastian Spicher, Christoph Bannwarth, and Stefan Grimme. A generally applicable atomic-charge dependent london dispersion correction. The Journal of chemical physics, 150(15), 2019.
  45. [46]DDMMS, STFC. ml-peg: ML performance and extrapolation guide. https://github.com/ddmms/ml-peg, 2025. Reference deployment: https://ml-peg.stfc.ac.uk.
  46. [47]James W. Furness, Aaron D. Kaplan, Jinliang Ning, John P. Perdew, and Jianwei Sun. Accurate and numerically efficient r²SCAN meta-generalized gradient approximation. The Journal of Physical Chemistry Letters, 11(19):8208–8215, 2020.
  47. [48]Eike Caldeweyher, Sebastian Ehlert, Andreas Hansen, Hagen Neugebauer, Sebastian Spicher, Christoph Bannwarth, and Stefan Grimme. A generally applicable atomic-charge dependent London dispersion correction. The Journal of Chemical Physics, 150(15):154122, 2019.
  48. [49]Sebastian Dohm, Andreas Hansen, Marc Steinmetz, Stefan Grimme, and Marek P. Checinski. Comprehensive thermochemical benchmark set of realistic closed-shell metal organic reactions. Journal of Chemical Theory and Computation, 14(5):2596–2608, 2018. doi: 10.1021/acs.jctc.7b01183.DOI
  49. [50]Sebastian Ehlert, Uwe Huniar, Jinliang Ning, James W. Furness, Jianwei Sun, Aaron D. Kaplan, John P. Perdew, and Jan Gerit Brandenburg. r²scan-d4: Dispersion corrected meta-generalized gradient approximation for general chemical applications. The Journal of Chemical Physics, 154(6):061101, 2021. doi: 10.1063/5.0041008.DOI
  50. [51]G. Kresse and J. Furthmüller. Efficient iterative schemes for ab initio total-energy calculations using a plane-wave basis set. Phys. Rev. B, 54:11169–11186, Oct 1996. doi: 10.1103/PhysRevB.54.11169. URL https://link.aps.org/doi/10.1103/PhysRevB.54.11169.DOI
  51. [52]G. Kresse and D. Joubert. From ultrasoft pseudopotentials to the projector augmented-wave method. Phys. Rev. B, 59:1758–1775, Jan 1999. doi: 10.1103/PhysRevB.59.1758. URL https://link.aps.org/doi/10.1103/PhysRevB.59.1758.DOI
  52. [53]G. Kresse and J. Furthmüller. Efficiency of ab-initio total energy calculations for metals and semiconductors using a plane-wave basis set. Comput. Mater. Sci., 6(1):15–50, 1996. ISSN 0927-0256. doi: https://doi.org/10.1016/0927-0256(96)00008-0. URL https://www.sciencedirect.com/science/article/pii/0927025696000080.
  53. [54]G. Kresse and J. Hafner. Ab initio molecular dynamics for liquid metals. Phys. Rev. B, 47:558–561, Jan 1993. doi: 10.1103/PhysRevB.47.558. URL https://link.aps.org/doi/10.1103/PhysRevB.47.558.DOI
  54. [55]Alex M Ganose, Hrushikesh Sahasrabuddhe, Mark Asta, Kevin Beck, Tathagata Biswas, Alexander Bonkowski, Joana Bustamante, Xin Chen, Yuan Chiang, Daryl C Chrzan, et al. Atomate2: Modular workflows for materials science. Digital discovery, 4(7):1944–1973, 2025.
  55. [56]A. S. Rosen, S. M. Iyer, D. Ray, Z. Yao, A. Aspuru-Guzik, L. Gagliardi, J. M. Notestein, and R. Q. Snurr. Machine learning the quantum-chemical properties of metal-organic frameworks for accelerated materials discovery. Matter, 4:1578–1597, 2021. doi: 10.1016/j.matt.2021.02.015.DOI
  56. [57]A. S. Rosen, V. Fung, P. Huck, C. T. O’Donnell, M. K. Horton, D. G. Truhlar, K. A. Persson, J. M. Notestein, and R. Q. Snurr. High-throughput predictions of metal–organic framework electronic properties: theoretical challenges, graph neural networks, and data exploration. npj Comput. Mater., 8:112, 2022. doi: 10.1038/s41524-022-00796-6.DOI
  57. [58]Andrew S. Rosen, Shaelyn M. Iyer, Debmalya Ray, Zhenpeng Yao, Alán Aspuru-Guzik, Laura Gagliardi, Justin M. Notestein, and Randall Q. Snurr. Machine learning the quantum-chemical properties of metal-organic frameworks for accelerated materials discovery. Matter, 4(5):1578–1597, 2021.
  58. [59]Theo Jaffrelot Inizan, Sherry Yang, Aaron Kaplan, Yen hsu Lin, Jian Yin, Saber Mirzaei, Mona Abdelgaid, Ali H. Alawadhi, KwangHwan Cho, Zhiling Zheng, Ekin Dogus Cubuk, Christian Borgs, Jennifer T. Chayes, Kristin A. Persson, and Omar M. Yaghi. System of agentic ai for the discovery of metal-organic frameworks, 2025. URL https://arxiv.org/abs/2504.14110.
  59. [60]Yongchul G. Chung et al. Advances, updates, and analytics for the computation-ready, experimental metal–organic framework database: CoRE MOF 2019. Journal of Chemical & Engineering Data, 64(12):5985–5998, 2019.
  60. [61]Marco Gibaldi, Anna Kapeliukha, Andrew White, Jun Luo, Robert Alex Mayo, Jake Burner, and Tom K Woo. Mosaec-db: a comprehensive database of experimental metal-organic frameworks with verified chemical accuracy suitable for molecular simulations. Chemical Science, 16(9):4085–4100, 2025.
  61. [62]Andrea Grisafi and Michele Ceriotti. Incorporating long-range physics in atomic-scale machine learning. The Journal of Chemical Physics, 151(20):204105, 2019.
  62. [63]Filip Formaliк, Alexander V. Neimark, Justyna Rogacka, Lucyna Firlej, and Bogdan Kuchta. Pore opening and breathing transitions in metal-organic frameworks: Coupling adsorption and deformation. Journal of Colloid and Interface Science, 578:77–88, 2020. ISSN 0021-9797. doi: https://doi.org/10.1016/j.jcis.2020.05.105. URL https://www.sciencedirect.com/science/article/pii/S0021979720307165.
  63. [64]Prathami Divakar Kamath and Kristin A Persson. The impact of spurious imaginary phonon modes on thermal properties of metal-organic frameworks. npj Computational Materials, 2026.
  64. [65]Wendy L Queen, Matthew R Hudson, Eric D Bloch, Jarad A Mason, Miguel I Gonzalez, Jason S Lee, David Gygi, Joshua D Howe, Kyuho Lee, Tamim A Darwish, et al. Comprehensive study of carbon dioxide adsorption in the metal–organic frameworks m2 (dobdc)(m= mg, mn, fe, co, ni, cu, zn). Chemical Science, 5(12):4569–4581, 2014.
  65. [66]Lukas Grajciar, Andrew D Wiersum, Philip L Llewellyn, Jong-San Chang, and Petr Nachtigall. Understanding co2 adsorption in cubic mof: comparing combined dft–ab initio calculations with microcalorimetry experiments. The Journal of Physical Chemistry C, 115(36):17925–17933, 2011.
  66. [67]Srinivasu Kancharlapalli and Randall Q Snurr. High-throughput screening of the core-mof-2019 database for co2 capture from wet flue gas: a multi-scale modeling strategy. ACS Applied Materials & Interfaces, 15(23):28084–28092, 2023.
  67. [68]Jongwoo Park, Joshua D Howe, and David S Sholl. How reproducible are isotherm measurements in metal–organic frameworks? Chemistry of Materials, 29(24):10487–10495, 2017.
  68. [69]Dominika O Wasik, José Manuel Vicent-Luna, Azahara Luna-Triguero, David Dubbeldam, Thijs JH Vlugt, and Sofía Calero. The impact of metal centers in the m-mof-74 series on carbon dioxide and hydrogen separation. Separation and Purification Technology, 339:126539, 2024.
  69. [70]Joshua Edzards, Julia Santana-Andreo, Holger-Dietrich Saßnick, and Caterina Cocchi. Benchmarking selected density functionals and dispersion corrections for mof-5 and its derivatives. Journal of Chemical Theory and Computation, 21(14):7062, 2025.
  70. [71]Karena W Chapman, Gregory J Halder, and Peter J Chupas. Guest-dependent high pressure phenomena in a nanoporous metal-organic framework material. Journal of the American Chemical Society, 130(32):10524–10526, 2008.
  71. [72]Pascal G Yot, Ke Yang, Florence Ragon, Vladimir Dmitriev, Thomas Devic, Patricia Horcajada, Christian Serre, and Guillaume Maurin. Exploration of the mechanical behavior of metal organic frameworks uio-66 (zr) and mil-125 (ti) and their nh2 functionalized versions. Dalton Transactions, 45(10):4283–4288, 2016.
  72. [73]Pascal G Yot, Zoubeyr Boudene, Jasmine Macia, Dominique Granier, Louis Vanduyfhuys, Toon Verstraelen, Veronique Van Speybroeck, Thomas Devic, Christian Serre, Gérard Férey, et al. Metal-organic frameworks as potential shock absorbers: the case of the highly flexible mil-53 (al). Chemical Communications, 50(67):9462–9464, 2014.
  73. [74]Louis R Redfern, Maxime Ducamp, Megan C Wasson, Lee Robison, Florencia A Son, François-Xavier Coudert, and Omar K Farha. Isolating the role of the node-linker bond in the compression of uio-66 metal-organic frameworks. Chemistry of Materials, 32(13):5864–5871, 2020.
  74. [75]Bin Zheng, Yihan Zhu, Fang Fu, Lian Li Wang, Jinlei Wang, and Huiling Du. Theoretical prediction of the mechanical properties of zeolitic imidazolate frameworks (zifs). RSC advances, 7(66):41499–41503, 2017.
  75. [76]Pia Vervoorts, Claire L Hobday, Michael G Ehrenreich, Dominik Daisenberger, and Gregor Kieslich. The zeolitic imidazolate framework zif-4 under low hydrostatic pressures. Zeitschrift für anorganische und allgemeine Chemie, 645(15):970–974, 2019.
  76. [77]F.A. Kloutse, R. Zacharia, D. Cossement, and R. Chahine. Specific heat capacities of mof-5, cu-btc, fe-btc, mof-177 and mil-53 (al) over wide temperature ranges: Measurements and application of empirical group contribution method. Microporous Mesoporous Mater., 217:1–5, 2015. ISSN 1387-1811. doi: https: //doi.org/10.1016/j.micromeso.2015.05.047. URL https://www.sciencedirect.com/ science/article/pii/S1387181115003248.
  77. [78]Hoa Thi Lai, Nhat Quang Minh Tran, Linh Ho Thuy Nguyen, Thu Bao Nguyen Le, Cuong Chi Nguyen, Anh Tuan Thanh Pham, Tan Le Hoang Doan, Sungkyun Park, Jongill Hong, Gerald Jeffrey Snyder, and Thang Bach Phan. Low experimental thermal conductivity of zirconium metal-organic framework uio-66. Appl. Phys. Lett., 124(15):152205, 04 2024. ISSN 0003-6951. doi: 10.1063/5.0201523. URL https://doi.org/10.1063/5.0201523.
  78. [79]Seyed Mohamad Moosavi, Balázs Álmos Novotny, Daniele Ongari, Elias Moubarak, Mehrdad Asgari, Özge Kadioglu, Charithea Charalambous, Andres Ortega-Guerrero, Amir H Farmahini, Lev Sarkisov, et al. A data-science approach to predict the heat capacity of nanoporous materials. Nat. Mater., 21(12):1419–1425, 2022.
  79. [80]Nina Lock, Mogens Christensen, Yue Wu, Vanessa K Peterson, Maja K Thomsen, Ross O Piltz, Anibal J Ramirez-Cuesta, Garry J McIntyre, Katarina Norén, Ramzi Kutteh, et al. Scrutinizing negative thermal expansion in mof-5 by scattering techniques and ab initio calculations. Dalton Trans., 42(6):1996–2007, 2013.
  80. [81]Yun Liu, Jae-Hyuk Her, Anne Dailly, Anibal J Ramirez-Cuesta, Dan A Neumann, and Craig M Brown. Reversible structural transition in mil-53 with large temperature hysteresis. Journal of the American Chemical Society, 130(35):11813–11818, 2008.
  81. [82]Matthew R Ryder, Bartolomeo Civalleri, Thomas D Bennett, Sebastian Henke, Svemir Rudić, Gianfelice Cinque, Felix Fernandez-Alonso, and Jin-Chong Tan. Identifying the role of terahertz vibrations in metal-organic frameworks: from gate-opening phenomenon to shear-driven structural destabilization. Physical review letters, 113(21):215502, 2014.
  82. [83]Matthew R Ryder, Ben Van de Voorde, Bartolomeo Civalleri, Thomas D Bennett, Sanghamitra Mukhopadhyay, Gianfelice Cinque, Felix Fernandez-Alonso, Dirk De Vos, Svemir Rudić, and Jin-Chong Tan. Detecting molecular rotational dynamics complementing the low-frequency terahertz vibrations in a zirconium-based metal-organic framework. Physical Review Letters, 118(25):255502, 2017.
  83. [84]Jian-Bin Lin, Tai TT Nguyen, Ramanathan Vaidhyanathan, Jake Burner, Jared M Taylor, Hana Durekova, Farid Akhtar, Roger K Mah, Omid Ghaffari-Nik, Stefan Marx, et al. A scalable metal-organic framework as a durable physisorbent for carbon dioxide capture. Science, 374(6574):1464–1469, 2021.
  84. [85]Jens Möllmer, Andreas Möller, Frieder Dreisbach, Roger Gläser, and Reiner Staudt. High pressure adsorption of hydrogen, nitrogen, carbon dioxide and methane on the metal–organic framework hkust-1. Microporous and Mesoporous Materials, 138(1-3):140–148, 2011.
  85. [86]Jason M Simmons, Hui Wu, Wei Zhou, and Taner Yildirim. Carbon capture in metal–organic frameworks—a comparative study. Energy & Environmental Science, 4(6):2177–2185, 2011.
  86. [87]Prashant Mishra, Hari Prasad Uppara, Bishnupada Mandal, and Sasidhar Gumma. Adsorption and separation of carbon dioxide using mil-53 (al) metal-organic framework. Industrial & Engineering Chemistry Research, 53(51):19747–19753, 2014.
  87. [88]Zhijuan Zhang, Shikai Xian, Qibin Xia, Haihui Wang, Zhong Li, and Jing Li. Enhancement of co2 adsorption and co2/n2 selectivity on zif-8 via postsynthetic modification. AIChE Journal, 59(6):2195–2206, 2013.
  88. [89]Hussein Rasool Abid, Huyong Tian, Ha-Ming Ang, Moses O Tade, Craig E Buckley, and Shaobin Wang. Nanosize zr-metal organic framework (uio-66) for hydrogen and carbon dioxide storage. Chemical Engineering Journal, 187:415–420, 2012.
  89. [90]Elliott Kasoar, Patrick Austin, Harvey Devereux, Kieran Harris, David Mason, Jacob Wilkins, Federica Zanca, and Alin Elena. janus-core.
  90. [91]Atsushi Togo, Laurent Chaput, Terumasa Tadano, and Isao Tanaka. Implementation strategies in phonopy and phono3py. J. Phys. Condens. Matter, 35(35):353001, 2023. doi: 10.1088/1361-648X/acd831.DOI
  91. [92]Atsushi Togo. First-principles phonon calculations with phonopy and phono3py. J. Phys. Soc. Jpn., 92(1):012001, 2023. doi: 10.7566/JPSJ.92.012001.DOI
  92. [93]Prathami Divakar Kamath, Francesco Tavani, Alin Marin Elena, Théo Jaffrelot Inizan, Yen-hsu Lin, Jian Yin, Wenqian Xu, Omar M Yaghi, and Kristin A Persson. Data-driven design of metal-organic frameworks with tunable negative thermal expansion. arXiv preprint arXiv:2607.18594, 2026.arxiv.org/abs/2607.18594

Paper details

vlnc:W_hhepbcme
Published
Added
License
CC BY 4.0
Usage

views · 0

downloads · 0

A view counts once per browser per day, a download each time the PDF is requested.

  • Physics and Astronomy

Not recorded for this paper

View statement
Looking for related papers…
Contents