{"id":"5f9a7850-fbda-41c2-8188-74724f844b3a","arxiv_id":"2509.06719","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":7,"one_line_summary":"A hybrid MOF screening workflow combining UFF and the PFP machine-learned potential, validated against DFT, identifies seven promising MOFs for humidity-tolerant ethylene capture.","lead":"A new screening pipeline grades over 1,800 MOFs first with a fast classical force field (UFF), then re-scores the top 88 with a universal machine-learned potential (PFP), benchmarking against DFT and accounting for framework flexibility. The workflow flags seven MOFs for selective ethylene capture under humid conditions and claims quantum-level accuracy at screening scale.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"PFP cell-relaxation predictions of guest-induced flexibility are not benchmarked against DFT; the headline 20 kJ/mol affinity shift and candidate ranking rest on unvalidated u-MLIP alone.","rationale":"The reader's concern about the PBE-D3 reference is valid but partially mitigated because the paper explicitly frames its predictions as DFT-level, not experimental. My concern targets the less-examined flexibility component: the PFP cell-relaxation results are not benchmarked against DFT at all. This is the load-bearing part of the central claim because the paper presents the up-to-20 kJ/mol deviations and the resulting candidate ranking (including the exclusion of some MOFs) as key outcomes. The internal benchmark in Figure 3 only covers rigid frameworks, so the flexibility results lack independent support. A DFT check on the extreme case QAQTEJ would settle whether the PFP structural response is reliable. The paper also provides code on GitHub and a transparent workflow, which are strengths, but they do not address this validation gap. Since the reader's verdict is already CONDITIONAL, my concern does not change the overall recommendation, but it sharpens the condition: flexibility predictions must be benchmarked before the claimed accuracy is accepted.","tokens_in":16677,"tokens_out":4328,"duration_ms":48371,"concrete_test":"Re-run the flexibility analysis for the outlier MOF QAQTEJ (CSD refcode: [from text]) using PBE-D3 DFT with full unit cell relaxation, using the same CP2K settings as in Section II.3 but allowing the cell to vary. Compare the optimized cell volume and ΔH0,ads(C2H4) with the PFP cell-opt values (cell volume contraction ~35%, ΔH0,ads increase from 31 to 51 kJ/mol). If DFT does not reproduce the ~20 kJ/mol affinity increase and the volume collapse, the flexibility claim and its impact on candidate selection are not supported.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim that the workflow accurately captures guest-induced framework flexibility is unsupported. The only PFP validation (Section II.3, Figure 3) benchmarks interaction energies for rigid frameworks (atomic positions relaxed, cell fixed) against PBE-D3 DFT, with MADs of 2.4 and 3.0 kJ/mol for C2H4 and H2O. However, the flexibility analysis in Section III ('Impact of MOF framework flexibility on ethylene affinity') uses PFP to fully relax unit cell parameters (configuration iii) and reports deviations in ΔH0,ads(C2H4) up to 20 kJ/mol, including a 35% volume contraction for QAQTEJ that raises affinity from 31 to 51 kJ/mol. This key result is never validated against DFT or experiment. Moreover, the geometries for the DFT benchmark are seeded by PFP's own Widom minima, so the benchmark may not expose PFP's errors in other configurations. Since the workflow's added value over UFF is precisely its ability to refine predictions for flexible MOFs, an unvalidated PFP cell-relaxation step could mis-rank candidates; the final list of seven MOFs might change if a different u-MLIP or DFT relaxation were used.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes a hierarchical screening workflow for MOF adsorbents: a UFF-based Widom-insertion first pass over a curated subset of the CSD MOF database, followed by re-evaluation of 88 top candidates with the PFP universal machine-learned interatomic potential, and a final DFT benchmark. The authors report PFP vs PBE-D3 interaction-energy MADs of 2.4 and 3.0 kJ/mol for C2H4 and H2O over 88 MOFs, identify seven MOFs with high ethylene affinity and C2H4/H2O selectivity, and analyze guest-induced framework flexibility, citing deviations in ethylene affinity up to 20 kJ/mol, including a 35% volume contraction for QAQTEJ.","tokens_in":17027,"tokens_out":5178,"duration_ms":46951,"significance":"If the claims hold, the workflow offers a practical template for combining classical force fields with u-MLIPs for large-scale MOF screening, and the paper provides a useful benchmark dataset and open code (https://github.com/gmaurin-group/MLP-WIDOM-SIM). The DFT validation across 88 chemically diverse MOFs is a genuine strength, and the identification of outlier cases where UFF fails (ZSTU-3, A520, Fe-CFA-6) gives concrete, falsifiable predictions. However, the most distinctive claim—that PFP reliably captures guest-induced flexibility—is not supported at the same level as the rigid-framework energetics, and the final candidate list depends on post hoc thresholds without sensitivity analysis.","major_comments":[{"comment":"The benchmark in Fig. 3 validates PFP only for rigid frameworks (atomic positions relaxed, cell fixed) and for interaction energies at PFP-selected Widom minima. The full cell-relaxation results in Fig. 5—including the 35% volume contraction and 20 kJ/mol affinity shift for QAQTEJ—are presented without any DFT or experimental check. Given that the value of the u-MLIP stage over UFF is argued to rest substantially on capturing flexibility, this is a load-bearing gap. Please add DFT relaxations (cell + atomic positions) for at least the outlier MOFs and several of the final candidates, or explicitly recast the flexibility discussion as a qualitative, hypothesis-generating result.","section":"§II.3, §III, Fig. 5"},{"comment":"The reference for benchmarking is PBE-D3 DFT, the same level that PFP is trained on. Agreement with PBE-D3 therefore does not establish agreement with experimental adsorption energetics. For a screening study that names seven specific sorbents, a comparison to experimental Henry constants or isotherms for at least a few well-known MOFs (e.g., ZIF-8, A520, or a cyclodextrin MOF) would calibrate the absolute accuracy. Without it, a systematic PBE-D3 bias for hydrogen-bonded or π-conjugated systems could affect the ranking of the final candidates.","section":"§II.3, Fig. 3"},{"comment":"The final affinity threshold (−ΔH0,ads(C2H4) > 43 kJ/mol) and selectivity threshold S(C2H4/H2O) > 50 are introduced after the u-MLIP results are displayed. There is no sensitivity analysis, and the number of 'seven top performers' is a direct consequence of these choices; small perturbations could add or remove candidates. Please report how many MOFs lie in the neighborhood of the thresholds and provide robustness checks (e.g., varying the affinity threshold by ±3 kJ/mol and S by a factor of 2).","section":"§III (Discussion of Figure 4f), Figure 4f"},{"comment":"The DFT benchmark geometries are seeded from the lowest-energy configurations found by PFP-based Widom insertion. This tests PFP's accuracy near its own minima but not its transferability to other guest configurations, which is exactly what the cell-relaxation and flexible-framework calculations require. The benchmark set therefore does not constrain the error of the flexibility results. Please state this limitation explicitly and, if possible, include a small set of DFT calculations starting from independent (e.g., random or UFF-derived) guest placements.","section":"§II.3 (PFP benchmarking paragraph)"}],"minor_comments":[{"comment":"Inconsistent naming: 'ZSTU-3' appears as 'ZTUS-3' in the text; 'ZSTU-380' and 'ZSTU-38080' are used interchangeably. Please standardize.","section":"Throughout"},{"comment":"The abstract and the discussion report the UFF vs PFP MAD for ΔH0,ads(C2H4) as both 5 and 5.1 kJ/mol. Use one value (5.1 kJ/mol) consistently.","section":"II.2 / III"},{"comment":"Reference 7 has a typo: 'Cundary, T. R.' should be 'Cundari, T. R.'; also 'Gordon, M. S.' is a co-author but the name order and initials should be checked against the original UFF paper.","section":"References"},{"comment":"The caption of Fig. 5 and the text discuss QAQTEJ and A520, but it would help to mark which subpanel shows each MOF and to define 'ΔH0,ads' at first use in the figure caption (it is defined in the methodology but not in the caption).","section":"Fig. 5 / III"},{"comment":"For the Widom insertion with PFP, only 50,000 MC cycles are reported. Please state the number of insertions per cycle or the statistical uncertainty of K_H and ΔH0,ads, as the UFF runs are described in more detail.","section":"II.2"}],"recommendation":"major_revision","confidential_remarks":"The manuscript is a solid application of an emerging tool, and the rigid-framework DFT benchmark is convincing. The main risk is the overreach in the flexibility conclusions, which are not validated at all. I would encourage the editor to insist on either DFT validation of the cell-relaxed structures or a clearly softened claim. The threshold analysis is a secondary but important robustness issue for the seven-candidate list."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague,\n\nThe useful thing about this paper is the workflow: UFF first pass on a curated set of ~1,900 neutral MOFs, then re-evaluation of 88 candidates with the PFP universal MLIP, with a DFT benchmark on those 88. The MADs for ethylene and water against PBE-D3 are 2.4 and 3.0 kJ/mol. That's a real result and the curation is careful. Code and data are on GitHub. This is a practical blueprint for hybrid screening.\n\nThe soft spot is the flexibility analysis. They use PFP to fully relax the unit cell in the presence of ethylene and report affinity changes up to 20 kJ/mol, including a 35% volume contraction for QAQTEJ. The DFT benchmark only covers rigid-framework geometries (atomic relaxation, fixed cell), so the flexible-cell predictions are never independently checked. The paper doesn't hide this, but it does lean on the flexibility step to exclude 'a few otherwise attractive MOFs.' That exclusion list is load-bearing and unvalidated. The silver lining: the final seven candidates all show small enthalpy changes under full relaxation (Table S1), so the top list is probably stable. Still, the authors should either run DFT on a couple of flexible cases or soften the claim.\n\nOther issues: the final thresholds (S>50, -ΔH>43 kJ/mol) appear to be selected after looking at the PFP results, so they are descriptive, not predictive. And the statement that the workflow works with 'any robust u-MLIP' is untested—only PFP was used. Also, 'near-DFT accuracy' is fair, but the abstract's phrasing might lead readers to expect experimental accuracy.\n\nBottom line: the method is sound, the benchmark is honest, and the seven candidates are testable. The problems are fixable in revision. I'd send it to peer review, with the main request being validation or calibration of the flexibility step, and a more careful framing of thresholds and transferability.","headline":"Hybrid UFF→u-MLIP screening with a solid 88-MOF DFT benchmark; flexibility claims outpace the validation.","tokens_in":17590,"tokens_out":4086,"would_cite":true,"duration_ms":42511,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"A hybrid screening workflow combining a classical force-field first pass with a universal machine-learned potential reaches near-DFT accuracy for ethylene/water adsorption in MOFs and identifies seven moisture-tolerant ethylene-selective ca","keywords":["metal-organic frameworks","high-throughput computational screening","machine-learned interatomic potentials","PFP","UFF","Widom insertion","ethylene/water selectivity","guest-induced flexibility"],"falsifier":"Compare the seven top-ranked MOFs' predicted low-coverage heats and Henry selectivities against experimental low-coverage adsorption calorimetry or isotherms measured under humid conditions; a systematic shortfall—several candidates below 43 kJ/mol or below selectivity 50—would falsify the workflow's ranking.","tokens_in":16531,"feed_emoji":"🧪","tokens_out":8269,"duration_ms":75500,"temperature":0.7,"pith_summary":"The paper tackles a practical bottleneck in computational materials screening: fast but approximate force fields can scan thousands of frameworks, yet they miss the subtle host–guest interactions that decide which material actually works. The authors propose a two-stage workflow using a generic classical force field, UFF, to preselect hydrophobic, ethylene-selective MOFs from a large database, then re-evaluate the 88 survivors with PFP, a universal machine-learned interatomic potential, benchmarking both against DFT. For most frameworks UFF and PFP agree, but for MOFs with hydrogen-bonding sites or tight confinement pockets UFF deviates by more than 10 kJ/mol, and PFP corrects those rankings. Including framework flexibility shifts ethylene affinity by up to about 20 kJ/mol for a subset of MOFs, so relaxing the unit cell matters. The final list contains seven MOFs with high ethylene affinity (more than 43 kJ/mol) and C2H4/H2O selectivity (greater than 50), proposed for humid ethylene-removal applications such as food packaging.","feed_headline":"Two-step screen finds seven MOFs for humid ethylene capture","feed_subtitle":"A cheap classical force field preselects candidates; a machine-learned potential re-checks them at near-DFT accuracy.","key_machinery":"The engine is a two-stage Widom-insertion Monte Carlo pipeline. In the first stage, UFF supplies Lennard-Jones parameters and MEPO-ML charges for rigid-framework test-particle insertions, producing Henry constants, infinite-dilution adsorption enthalpies, and ideal selectivity for both ethylene and water. In the second stage, PFP—the PreFerred Potential, an equivariant graph-neural-network interatomic potential trained on a large DFT-derived dataset—evaluates host–guest energies for the same Widom insertions, with a PBE-D3 dispersion correction applied. The second stage also relaxes atomic positions and unit-cell parameters under a single adsorbed guest, converting the screening from rigid-f","core_discovery":"The central claim is that accuracy and scalability in MOF adsorption screening need not be traded off: a generic classical force field (UFF) can serve as a cheap first-pass filter over thousands of frameworks, and a universal machine-learned potential (PFP u-MLIP), validated against PBE-D3 DFT, can refine the top candidates. Comparing UFF and PFP on 88 MOFs shows UFF captures the qualitative hydrophilic/hydrophobic ranking and ethylene affinities within a mean absolute deviation of about 5 kJ/mol, but a small set of outliers—MOFs with μ-OH groups in V-shaped pockets and one water-binding outlier—deviate beyond 10 kJ/mol; PFP captures the short π···OH contacts and confined water arrangements","pith_inferences":["Because the PFP benchmark is against DFT rather than experiment, the seven candidates' ranking should be checked against measured isotherms; a mismatch would localize the error to the DFT reference rather than to the workflow itself.","The observation that unit-cell relaxation changes ethylene affinity by up to 20 kJ/mol implies that other rigid-framework high-throughput screening studies may have systematically mis-ranked flexible MOFs; re-ranking a full database with cell relaxation is a natural next test.","The same triage logic—cheap generic model first, more expensive learned potential on the shortlist—could extend to predicting diffusivities, open-metal-site binding, or multicomponent co-adsorption, not just single-component heats.","The paper reports only ideal, Henry-regime selectivity; real humid-mixture selectivity could differ, so multicomponent grand-canonical Monte Carlo with PFP on the seven finalists is a direct testable extension."],"forward_implications":["UFF-based rigid-framework screening remains a defensible first pass for roughly 97–99% of the MOFs tested; only a small minority need higher-fidelity re-evaluation.","For MOFs with hydrogen-bonding groups, narrow pores, or confinement pockets, a universal machine-learned potential changes adsorption geometries and energies enough to alter rankings, so the hybrid stage is necessary for identifying the final top candidates.","Guest-induced framework flexibility, especially unit-cell relaxation, can change ethylene affinity by up to about 20 kJ/mol; ignoring it can mis-rank flexible MOFs such as layered structures.","The seven identified MOFs—with pore sizes around 4.5–6.0 Å, ethylene affinity greater than 43 kJ/mol, and C2H4/H2O selectivity greater than 50—are concrete targets for humid-condition ethylene-removal applications.","The same two-stage strategy can be applied to other adsorption separations and to databases of more than 100,000 structures, without being tied to a specific machine-learned potential."],"supporting_citations":[{"why":"Supplies the Lennard-Jones parameters for the UFF first-pass screen of MOF–guest interactions.","marker":"7"},{"why":"Defines PFP, the universal machine-learned potential that carries the refinement stage and matches DFT on the benchmark set.","marker":"56"},{"why":"Grimme D3 dispersion correction used in both PFP and DFT energy evaluations, enabling consistent treatment of van der Waals interactions.","marker":"69"},{"why":"CP2K/Quickstep DFT calculations provide the PBE-D3 reference energies against which PFP is benchmarked.","marker":"73"},{"why":"Widom test-particle insertion is the statistical-mechanical method used to obtain Henry constants and infinite-dilution adsorption enthalpies.","marker":"63"},{"why":"MEPO-ML assigns atomic partial charges to MOF atoms in the UFF simulations, controlling electrostatics in the first-pass screen.","marker":"65"},{"why":"Supplies the united-atom ethylene guest model used in the classical simulations.","marker":"66"},{"why":"Supplies the TIP4P-Ew water guest model used in the classical simulations.","marker":"67"},{"why":"Provides the CSD MOF database from which the 88 candidates were curated and screened.","marker":"55"}],"fun_headline_variants":["Hybrid screen finds 7 MOFs for humid ethylene capture","UFF plus ML potential picks 7 humid-stable MOFs","7 MOFs found by two-tier screening for ethylene","Machine-learned potential refines MOF adsorption screen","Seven MOFs pass hybrid screening for humid ethylene"],"cache_read_input_tokens":2688,"weakest_assumption_plain":"The screening treats one particular density-functional-theory calculation (PBE-D3) as the correct answer for how strongly ethylene and water bind in every framework; if that reference is biased, the final candidates could be mis-ranked.","fun_headline_variants_meta":{"raw":{"variants":["Hybrid screen finds 7 MOFs for humid ethylene capture","UFF plus ML potential picks 7 humid-stable MOFs","7 MOFs found by two-tier screening for ethylene","Machine-learned potential refines MOF adsorption screen","Seven MOFs pass hybrid screening for humid ethylene"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000184,"raw_usage":{"total_tokens":1229,"prompt_tokens":893,"completion_tokens":336,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":637,"completion_tokens_details":{"reasoning_tokens":256}},"tokens_in":637,"tokens_out":336,"duration_ms":4033,"temperature":1.0,"reasoning_tokens":256,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-04T23:11:31.084423+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Compare the seven top-ranked MOFs' predicted low-coverage heats and Henry selectivities against experimental low-coverage adsorption calorimetry or isotherms measured under humid conditions; a systematic shortfall—several candidates below 43 kJ/mol or below selectivity 50—would falsify the workflow's ranking.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Defines PFP, the universal machine-learned potential that carries the refinement stage and matches DFT on the benchmark set."},{"cited_title":"& Krieg, H","cited_arxiv_id":null,"evidence_quote":"Grimme D3 dispersion correction used in both PFP and DFT energy evaluations, enabling consistent treatment of van der Waals interactions."},{"cited_title":"& VandeVondele, J","cited_arxiv_id":null,"evidence_quote":"CP2K/Quickstep DFT calculations provide the PBE-D3 reference energies against which PFP is benchmarked."},{"cited_title":"& Smit, B","cited_arxiv_id":null,"evidence_quote":"Widom test-particle insertion is the statistical-mechanical method used to obtain Henry constants and infinite-dilution adsorption enthalpies."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"MEPO-ML assigns atomic partial charges to MOF atoms in the UFF simulations, controlling electrostatics in the first-pass screen."},{"cited_title":"& Fröba, M","cited_arxiv_id":null,"evidence_quote":"Supplies the united-atom ethylene guest model used in the classical simulations."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the TIP4P-Ew water guest model used in the classical simulations."},{"cited_title":"R., Bruno, I","cited_arxiv_id":null,"evidence_quote":"Provides the CSD MOF database from which the 88 candidates were curated and screened."}],"review_version":1}