REVIEW 3 major objections 6 minor 2 references
Decoding local framework dynamics in the ultra-small pore MOF MIL-120(Al) CO2 sorbent with Machine Learned Potentials
T0 review · 3 major / 6 minor · reviewed 2026-08-15 · deepseek-v4-flash
Pith's one-line read Hidden hydroxyl orientations set how CO2 binds in MIL-120(Al).
desk verdict Solid computational study showing μ2-OH orientation is a hidden degree of freedom in MIL-120(Al); the PBE-D3-only energy ranking is the main caveat, not the MLP circularity. read the letter →
The pith
A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.
The reading
What carries the argument
The load-bearing object is the orientation pattern of the μ2-OH bridging hydroxyl groups in the MIL-120(Al) unit cell; each distinct pattern defines one framework configuration, and the paper treats this pattern as a hidden structural degree of freedom. The machine-learned potential carries the computation: trained on roughly 183,000 DFT-derived snapshots, including transition-path images, strained structures, and CO2-loaded configurations, it evaluates energies and forces with near-DFT accuracy, making it feasible to map the full interconversion landscape and to scan CO2 adsorption sites without the cost of direct DFT at every step. The hydrogen-bond network formed in the lowest-energy configuration A is the microscopic mechanism that stabilises that state, and the transient breaking and re-forming of these hydrogen bonds is the mechanism behind the low barriers.
What would settle it
A direct measurement of hydrogen positions by neutron powder diffraction or 2H solid-state NMR on MIL-120(Al) at room temperature would settle the ground-state ordering: if the dominant hydroxyl orientation is not the interlocking hydrogen-bond pattern of configuration A, or if hydroxyl flipping is not observed on experimental timescales, the claimed energy ranking and dynamic accessibility would be wrong.
Extended reading notes
Core claim
The central claim is that the local dynamics of the μ2-OH groups in MIL-120(Al) form a key, previously overlooked feature controlling CO2 capture. Concretely, the paper constructs six MIL-120(Al) models that differ only in the orientation of the four μ2-OH groups per unit cell. After DFT geometry optimization, MIL-120(Al)-A is the most stable and MIL-120(Al)-F, the model used in earlier work, is the least stable, with an energy gap of about 0.59 eV per unit cell. Transition-state calculations give interconversion barriers of 0.07–0.19 eV per unit cell, implying all six states are dynamically accessible at room temperature, and phonon calculations show each is dynamically stable. When CO2 is introduced, its molecular axis aligns perpendicular to the one-dimensional pore when μ2-OH groups point into the channel, and parallel when the hydroxyls are positioned more axially; this alignment difference is what sets the adsorption energetics. The paper's machine-learned potential reproduces the DFT energies, barriers, phonons, and CO2 adsorption geometries within a few kJ mol$^{-1}$, which is what allows the full hydroxyl configurational landscape to be mapped.
Load-bearing premise
The energy ordering and the 0.07–0.19 eV barrier window rest on a single density-functional approximation for hydrogen-bond strengths and on the assumption that the six hand-selected μ2-OH orientation patterns are representative, an assumption X-ray diffraction cannot check because it does not locate hydrogen atoms.
Editorial extensions
If this is right
- At room temperature, MIL-120(Al) should be treated as a dynamic mixture of at least six μ2-OH configurations; any single static structure is an idealisation.
- The commonly used MIL-120(Al)-F model overestimates the isosteric heat of CO2 adsorption by about 27% relative to experiment, while the A, B, and C configurations match it.
- CO2 adsorption geometry follows the local hydroxyl orientation: perpendicular alignment when OH points into the channel, parallel alignment when OH lies axial to the channel.
- CO2 does not lock the framework: its presence changes the hydroxyl reorientation barriers by only a few percent, and in some pathways slightly lowers them.
- Generic pre-trained machine-learned potentials can systematically over- or under-estimate CO2 interaction energies in this polar, flexible MOF; system-specific training is needed for reliable predictions.
Reading between the lines
- The same hidden-degree-of-freedom mechanism is likely at play in other ultra-small-pore MOFs whose pores are decorated with bridging hydroxyls, so structural models that fix hydrogen positions may need revisiting there too. A direct test would be low-temperature neutron diffraction or 2H NMR that can locate H atoms.
- If configuration A really is the ground state, earlier computational studies that adopted F as the representative structure may have systematically biased their reported CO2 capacities and selectivities; re-running those calculations with A could change rankings among Al-MOF sorbents.
- The low, guest-insensitive interconversion barriers suggest MIL-120(Al) could show measurable hydroxyl reorientation dynamics in variable-temperature infrared or inelastic neutron spectroscopy, which would be a direct experimental fingerprint of the proposed flipping.
- The apparent success of a system-specific MLP in reproducing transition states hints that the strategy of training on CI-NEB images plus AIMD snapshots could be ported to other flexible MOFs with functional-group rotations, where generic MLIPs are weakest.
Signed reviews
Editorial analysis
A structured set of objections, weighed in public.
Referee Report
Summary. The paper combines PBE-D3 density-functional theory (DFT) with a purpose-trained DeePMD machine-learned potential (MLP) to investigate the orientations of bridging μ2-OH groups in the ultra-small pore MOF MIL-120(Al). Six manually constructed configurations (A–F) differing only in μ2-OH orientation are optimized, and the authors report that configuration A is the lowest-energy form while the commonly used F configuration is the highest, with an energy difference of about 0.59 eV per unit cell. CI-NEB calculations give interconversion barriers of 0.07–0.19 eV per unit cell, implying all six configurations are dynamically accessible at room temperature. The trained MLP is shown to reproduce DFT energies, barriers, phonon spectra, and CO2 adsorption geometries, and Widom insertion with the MLP yields isosteric heats of adsorption that match experiment for configurations A, B, and C. The central claim is that μ2-OH orientation is a 'hidden' structural degree of freedom controlling pore size and CO2 binding geometry (parallel vs. perpendicular), which has been overlooked in previous rigid-framework studies.
Significance. If the main conclusions hold, the paper provides a valuable correction to the standard structural model of MIL-120(Al) (F vs. A) and demonstrates that local hydroxyl dynamics can decisively control guest location and energetics in ultra-small pore MOFs. The work combines direct DFT energy calculations, transition-state searches, phonon analysis, and a large, openly available MLP training set (183,061 snapshots; data on GitHub), which is a reproducible and transparent computational resource. The identification of a hidden degree of freedom that is invisible to X-ray diffraction is conceptually important for the broader MOF adsorption community, where rigid-framework and fixed-hydrogen models are common. The MLP's near-DFT-level reproduction of energetics and geometries, while partly expected from the training set composition, still enables extensive sampling (Widom insertion, large-scale MD), and the direct DFT results provide a solid base for the central physics claim.
major comments (3)
- [Methods (DFT calculations) and Results (Exploration)] The relative energies of the six configurations and the 0.07–0.19 eV interconversion barriers are computed exclusively with PBE-D3 (Methods, 'DFT calculations'). The configurations differ chiefly by an interlocking hydrogen-bond network among μ2-OH groups, and PBE-D3 is known to be delicate for hydrogen-bond energetics: PBE tends to under-bind H-bonds while D3 dispersion can over-stabilize compact arrangements. Because the claim that A is the ground state and that all six states are thermally accessible rests directly on this energy ordering, the authors should test at least one additional functional (e.g., r2SCAN, SCAN-D3, or a hybrid) for the empty-framework relative energies and for the lowest-energy interconversion barrier, or provide a targeted justification why PBE-D3 is reliable for this specific H-bond network. Without such a check, the central energetic conclusion is not robust to functional choice.
- [Methods (Dataset preparation for MLP training)] The MLP training set explicitly includes CI-NEB intermediate images, AIMD snapshots, and CO2-loaded configurations (Methods, 'Dataset preparation for MLP training'). Therefore, the reported MLP reproduction of the DFT CI-NEB barriers, phonon spectra, and CO2 adsorption geometries is an interpolation check on the training set rather than an independent prediction. The text states that the MLP 'quantitatively reproduced the DFT-derived barriers and transition-state geometries' and 'accurately predicting CO2 adsorption geometries', which overstates the evidential weight of these comparisons. The authors should explicitly acknowledge this circularity and, if possible, validate the MLP on a transition pathway or configuration held out from training (e.g., a leave-one-out CI-NEB path or a new AIMD trajectory at a temperature not sampled), or temper the predictive claims accordingly.
- [Results (Exploration of the local structural features)] The six configurations are hand-selected 'representative' structures, and no enumeration, symmetry analysis, or systematic search is presented to show that they cover the full landscape of μ2-OH orientations. The paper later refers to 'all these states can be observed' and implies that the six configurations span the relevant structural space. If a seventh low-energy orientation pattern exists or the true ground-state pattern differs from A, the reported Qst match for A/B/C would be coincidental rather than diagnostic. The authors should provide a systematic search (e.g., enumeration of symmetry-inequivalent orientation patterns or Monte Carlo/temperature-accelerated sampling) to support the representativeness of the six configurations, or explicitly state in the Discussion that the six patterns are illustrative and not proven exhaustive.
minor comments (6)
- [Abstract] The phrase 'throughout reorientation of functional groups' should read 'through reorientation of functional groups' (typo).
- [Methods (DFT calculations)] The text refers to 'Supplementary Nate2'; this should be 'Supplementary Note 2'.
- [Fig. 4b and main text] The MLP name is given as 'MACE-MPA-0' in the figure caption and 'MACE-MP-0' in the main text; the notation should be made consistent.
- [Fig. 4 caption] The interaction-energy formula is corrupted in the extracted text (e.g., '𝐸!"#=𝐸$%!@’%(-(𝐸’%+𝐸$%!)'); please provide a clean, standard equation for E_int.
- [Results (Machine-learned potential development)] The statement that the MLP achieves 'near-DFT-level fidelity' should be qualified, because the RMSE values (0.217 and 0.268 meV/atom) are computed on training or closely related configurations rather than on an independent test set; a brief clarification would help.
- [Discussion] The phrase 'a large set of configurations' is used while only six were constructed; consider rewording to 'a set of six representative configurations' to avoid implying a more exhaustive sampling than was performed.
Circularity Check
The MLP's 'reproduction' of DFT barriers is interpolation over CI-NEB images already included in its training set; the central DFT-NEB energy landscape and the CO2-orientation correlation rest on direct DFT and are not circular.
-
fitted input called prediction
[Results, subsection 'Structural transitions and associated local μ2-OH reorientations between MIL-120(Al) configurations predicted by DFT and MLP calculations'; Methods, 'Dataset preparation for MLP…]
"Importantly, the MLP-predicted pathways quantitatively reproduced the DFT-derived barriers and transition-state geometries across all computed systems."
The Methods state that the MLP training dataset includes 'CI-NEB intermediate images sampled along μ2-OH reorientation paths' and AIMD snapshots. The MLP's agreement with the DFT barriers is therefore interpolation over explicitly sampled transition-state geometries, not an independent prediction of those barriers. The paper presents this agreement as notable validation, but the match is by construction because the transition states are part of the training set. This does not make the underlying DFT-NEB barriers circular, since those are computed directly at the DFT level; it only removes the MLP agreement as independent evidence for the barrier heights.
full rationale
The paper's main derivation chain is: (1) six MIL-120(Al) configurations with distinct μ2-OH orientations are DFT-optimized, giving A as the lowest-energy form and F as the highest (ΔE ≈ 0.59 eV/uc); (2) DFT-CI-NEB gives interconversion barriers of 0.07–0.19 eV/uc; (3) DFT and MLP relaxations of CO2-loaded structures correlate CO2 orientation with μ2-OH orientation; (4) MLP-based Widom Qst values for A/B/C match experiment. Steps (1)–(3) are supported by direct PBE-D3 DFT calculations, not by the MLP. The MLP is a fitted surrogate trained on a MIL-120-specific DFT dataset, so statements such as 'the MLP-predicted pathways quantitatively reproduced the DFT-derived barriers' describe interpolation, not independent confirmation; this is the one clear fitted-input-called-prediction issue. It does not infect the central claims because the DFT-NEB barriers and DFT-optimized CO2 geometries stand on their own. The hand-selection of six configurations and the absence of a second DFT functional are correctness risks, but not circularity: the paper never claims these configurations are exhaustive, and a functional sensitivity test is absent rather than assumed. No load-bearing self-citation or imported uniqueness theorem was found. Hence a score of 4 reflects one supporting 'prediction' that reduces by construction, while the central result retains independent DFT content.
Assumptions & free parameters
free parameters (3)
- DeePMD radial cutoff =
7.8 Angstrom
- DeePMD embedding and fitting network sizes =
Embedding {25,50,100}, fitting {240,240,240}
- MLP training weights =
Not published (trained on 183,061 DFT snapshots)
assumptions (4)
- domain assumption PBE-D3 exchange-correlation functional accurately describes hydrogen bonding and dispersion in this framework.
- ad hoc to paper The six manually constructed mu2-OH orientation models are representative of the full accessible configurational landscape.
- domain assumption The MLP training set, which includes CI-NEB transition images and AIMD snapshots, adequately covers the regions used to validate the MLP.
- domain assumption DFT-relaxed H positions correspond to physical H positions even though X-ray diffraction cannot locate them.
Cite this review
Pith. "Pith review of Decoding local framework dynamics in the ultra-small pore MOF MIL-120(Al) CO2 sorbent with Machine Learned Potentials." pith.science (2026). https://pith.science/paper/ALLTM2JF
@misc{pith2026250820608,
author = {Pith},
title = {Pith review of: Decoding local framework dynamics in the ultra-small pore MOF MIL-120(Al) CO2 sorbent with Machine Learned Potentials},
year = {2026},
howpublished = {\url{https://pith.science/paper/ALLTM2JF}},
note = {Machine review of arXiv:2508.20608}
}
read the original abstract
Metal-organic frameworks (MOFs) with ultra-small pores offer an optimal environment to effectively capture guest molecules such as CO2. Subtle local dynamics of their frameworks, either throughout reorientation of functional groups grafted to the organic linkers or those present in their inorganic nodes, is expected to play a major role in their sorption behaviors. Here, we combine density-functional theory (DFT) with a purpose-trained machine-learned potential to systematically investigate the local dynamics of the bridging hydroxyl groups, {\mu}2-OH groups present in the prototypical ultra-small pore MOF MIL-120(Al), reported recently as an attractive CO2 sorbent. We identified six MOF configurations associated with distinct {\mu}2-OH orientations with relatively low interconversion energy barriers (0.07-0.19 eV per unit cell) suggesting that all these states can be observed experimentally at room temperature. We demonstrated that our MLP achieves near-DFT-level fidelity, reproducing the energy barriers and phonon spectra of the empty MOF, and accurately predicting CO2 adsorption geometries depending on the {\mu}2-OH orientations with CO2 adopting either parallel or perpendicular alignment to the pore axis, which in turn governs the adsorption energetics. This work establishes that a reliable description of the local structure, such as reorientation/flipping of bridging hydroxyl groups, is a key feature to gain an accurate description of the guest locations and energetics in ultra-small pore MOFs.
Figures
Reference graph
Works this paper leans on
-
[15]
Chung, Y. G. et al. Advances, Updates, and Analytics for the Computation-Ready, Experimental Metal–Organic Framework Database: CoRE MOF 2019. J. Chem. Eng. Data 64, 5985–5998 (2019). 16. Volkringer, C. et al. Occurrence of Uncommon Infinite Chains Consisting of Edge-Sharing Octahedra in a Porous Metal Organic Framework-Type Aluminum Pyromellitate Al4 (OH)...
-
[29]
Blöchl, P. E. Projector augmented-wave method. Phys. Rev. B 50, 17953–17979 (1994). 30. Perdew, J. P., Burke, K. & Ernzerhof, M. Generalized Gradient Approximation Made Simple. Phys. Rev. Lett. 77, 3865–3868 (1996). 31. Monkhorst, H. J. & Pack, J. D. Special points for Brillouin-zone integrations. Phys. Rev. B 13, 5188–5192 (1976). 32. Grimme, S. Semiempi...
work page 1994
Reviewed August 15, 2026 · model on record in the stance chip above.
Discussion (0). Continue with ORCID to comment.